Opened 7 years ago

Closed 7 years ago

#541 closed Request (fixed)

net.sf.basedb.normalizers: Normalizers should store normalized data in proper base

Reported by: Jari Häkkinen Owned by: olle
Priority: major Milestone: Normalization package v1.1
Component: net.sf.basedb.normalizers Keywords:
Cc:

Description

QuantileNormalizer always stores data non-logged irrespective what data was before normalization. Data should be stored back to BASE in the same base as before the transform.

Change History (14)

comment:1 Changed 7 years ago by olle

Status: newassigned

Ticket accepted.

comment:2 Changed 7 years ago by olle

Background discussion (Thanks to Nicklas Nordborg for detailed information about this):

  1. Originally data in BASE was stored in untransformed format, and the average method used was arithmetic mean (the latter differs from the current recommendation for untransformed data, which is to use geometric mean).
  2. A flag indicating if stored data was untransformed, stored as log-2, or log-10 values, was introduced in BASE Ticket #1120 (The dynamic part of BASE should keep track whether intensity data is in log space or not). See this ticket for a lengthy discussion on the use and storage of transformed data in BASE. Some additions were:

    a. A BioAssaySet has a getIntensityTransform() method that returns an IntensityTransform enum object, whose public Formula.AverageMethod getAverage() method returns Formula.AverageMethod.ARITHMETIC_MEAN or Formula.AverageMethod.GEOMETRIC_MEAN.
    b. Class VirtualColumn was extended with static methods VirtualColumn channelIntensity(int channel) and VirtualColumn channelRaw(int channel). The former performs a reverse transformation if needed, determined by the transformation flag, in order to return untransformed data, while the latter returns data the way it was stored in the database, i.e. data in log-2 and log-10 format is returned in that format.

  3. Enum class IntensityTransform public Formula.AverageMethod getAverage() method does however return Formula.AverageMethod.ARITHMETIC_MEAN for data stored in untransformed format (according to the flag), and Formula.AverageMethod.GEOMETRIC_MEAN for data stored in log-2 or log-10 format. This is the opposite to what is recommended to be used in this ticket, for data fetched in untransformed form.
Last edited 7 years ago by olle (previous) (diff)

comment:3 Changed 7 years ago by olle

Traceability note:

  • BASE Ticket #1120 (The dynamic part of BASE should keep track whether intensity data is in log space or not) added several new methods and items concerning work with transformed/untransformed data.
  • Ticket #228 (net.sf.basedb.normalizers: Normalizers should use ask BASE whether data is logged or not) is concerned with setting a proper default average method for normalization, based on if the data is stored in transformed or untransformed format.

comment:4 Changed 7 years ago by olle

Problem discussion:

  • Class QuantileNormalization in src/net/sf/basedb/plugins/QuantileNormalization.java extends AbstractNormalizationPlugin, and the latter contains several methods to access data, all using method VirtualColumn.channelIntensity(int channel) to retrieve the data from the database. For transformed data in log-2 or log-10 format, this means that the data is returned in untransformed format, and will be stored as such after normalization.
  • Exchanging method VirtualColumn.channelIntensity(int channel) calls in class AbstractNormalizationPlugin for VirtualColumn.channelRaw(int channel) calls, will result in the data being returned as stored in the database, without any optional transformation, and therefore will be stored after normalization in the same format as when fetched. For this to work, it is essential that a proper averaging method is used, as using geometric mean for data with zero or negative values leads to problems.
Last edited 7 years ago by olle (previous) (diff)

comment:5 Changed 7 years ago by olle

Problem discussion update:

  • BASE Ticket #1792 (Incorrect average method specified in IntensityTransform) switched the results returned by public method Formula.AverageMethod getAverage() in enum class IntensityTransform for data flagged to be stored in untransformed, relative to transformed (log-2 or log-10) format (see changeset [6564]):

    Formula.AverageMethod.GEOMETRIC_MEAN for data stored in untransformed format
    Formula.AverageMethod.ARITHMETIC_MEAN for data stored in transformed format

    This would solve the problem with wrong default averaging method being shown for QuantileNormalizer. However, there are benefits letting the plug-in itself determine the default averaging method, based on the intensity transform information for the BioAssaySet to work on.

    Class IntensityTransform was extended with new public method double transform(double value) in changeset [6365], that stores data in the same transform as the source data.

comment:6 Changed 7 years ago by olle

Design update for QuantileNormalization:

  1. Data stored in logarithmic format should be untransformed before averaging, and then transformed back to logarithmic format before results are stored. The first part is already implemented by use of static method VirtualColumn channelIntensity(int channel) in class AbstractNormalizationPlugin, while public method double transform(double value) in class IntensityTransform should be used to transform the normalized result back before storing results.
  2. Since averaging is performed on untransformed data, default averaging method should always be set to Formula.AverageMethod.GEOMETRIC_MEAN.
  3. Help text for selecting averaging method in class AbstractNormalizationPlugin should be updated to avoid references to the format data is stored in, since averaging is performed on untransformed data.
  4. Help text for Quantile normalization in META-INF/extensions.xml should be extended with information that data stored in logarithmic format will be untransformed before averaging, and then transformed back to logarithmic format before results are stored.
Last edited 7 years ago by olle (previous) (diff)

comment:7 Changed 7 years ago by olle

(In [2165]) Refs #228. Refs #541. Quantile normalization updated to untransform logarithmic data before normalization, and then transform it back to logarithmic format before results are stored. Default averaging method will always be set to Formula.AverageMethod.GEOMETRIC_MEAN:

  1. Class/file QuantileNormalizer.java in src/net/sf/basedb/plugins/ in package net.sf.basedb.normalizers updates:
    a. Private method RequestInformation getConfiguredJobParameters() updated to always set the default averaging method to Formula.AverageMethod.GEOMETRIC_MEAN.
    b. Private method BioAssaySet normalize(DbControl dc, BioAssaySet source, Job job, ProgressReporter progress) updated to call public method double transform(double value) in class IntensityTransform to transform the normalized result back before storing it. Error message when number of spots differ between two BioAssay sets updated to report the names of the latter and the number of spots in each one. Also minor updates in order to increase clarity of code.
  2. Class/file AbstractNormalizationPlugin.java in src/net/sf/basedb/plugins/ in package net.sf.basedb.normalizers updated in help text for selecting averaging method by avoiding reference to the format data is stored in, since averaging is performed on untransformed data.
  3. XML files extensions.xml in META-INF in package net.sf.basedb.normalizers updated in help text for Quantile normalization by adding information that data stored in logarithmic format will be untransformed before averaging, and then transformed back to logarithmic format before results are stored.

comment:8 Changed 7 years ago by olle

Design update for AverageNormalization (mirrors update for QuantileNormalization):

  1. Data stored in logarithmic format should be untransformed before averaging, and then transformed back to logarithmic format before results are stored. The first part is already implemented by use of static method VirtualColumn channelIntensity(int channel) in class AbstractNormalizationPlugin, while public method double transform(double value) in class IntensityTransform should be used to transform the normalized result back before storing results.
  2. Since averaging is performed on untransformed data, default averaging method should always be set to Formula.AverageMethod.GEOMETRIC_MEAN.

comment:9 Changed 7 years ago by olle

(In [2166]) Refs #228. Refs #541. Average normalization updated to untransform logarithmic data before normalization, and then transform it back to logarithmic format before results are stored. Default averaging method will always be set to Formula.AverageMethod.GEOMETRIC_MEAN:

  1. Class/file AverageNormalization.java in src/net/sf/basedb/plugins/ in package net.sf.basedb.normalizers updates:
    a. Private method RequestInformation getConfiguredJobParameters() updated to always set the default averaging method to Formula.AverageMethod.GEOMETRIC_MEAN.
    b. Private method BioAssaySet normalize(DbControl dc, BioAssaySet source, Job job, float refValue, float minIntensity, ProgressReporter progress) updated to call public method double transform(double value) in class IntensityTransform to transform the normalized result back before storing it.

comment:10 Changed 7 years ago by olle

Design note:

  • When data stored in logarithmic format are untransformed before averaging, and then transformed back to logarithmic format before results are stored, transformation information for the result BioAssaySet must be set explicitly by calling its public void setIntensityTransform(IntensityTransform transform) method. If this is not done, methods using the result BioAssaySet will treat the data in logarithmic format as original data.

comment:11 Changed 7 years ago by olle

(In [2167]) Refs #541. Rank invariant normalization updated to store data in same format as before the normalization, untransformed, log-2, log-10. (Note that Rank inveriant normalization currently is only implemented for one-channel data):

  1. Class/file RankInvariantNormalization.java in src/net/sf/basedb/plugins/ in package net.sf.basedb.normalizers updated in private method BioAssaySet normalize1Ch(DbControl dc, BioAssaySet source, List<?> masterAssays, int numIteration, Job job, ProgressReporter progress) to call public method double transform(double value) in class IntensityTransform to transform the normalized result back before storing it.
Last edited 7 years ago by olle (previous) (diff)

comment:12 Changed 7 years ago by Jari Häkkinen

(In [2173]) Refs #228 and #541. Improving doc and help string

comment:13 Changed 7 years ago by olle

(In [2174]) Refs #541. Refs #228. Outermost Ant build XML file updated by increasing dependency base-version from 3.0.0 to 3.2.4, as this is required to use stored information on input data format (pristine, log-2, or log-10).

comment:14 Changed 7 years ago by olle

Resolution: fixed
Status: assignedclosed

Ticket closed as the requested functionality has been added.

Note: See TracTickets for help on using tickets.