# evigene/docs/evigene-tr2aacds-classifier.txt
# don gilbert, 2014 

Classification versus Clustering and EvidentialGene tr2aacds mRNA transcript classifier

Classification: separating data into categories, or classes, characterized by a distinct set of 
features or measures.  An algorithm that implements classification, especially in a concrete 
implementation, is known as a classifier.

Clustering: Grouping a set of objects based on measures, so that objects in the same group
are more similar to each other than to those in other groups.

Eg. http://en.wikipedia.org/wiki/Statistical_classification
    http://en.wikipedia.org/wiki/Cluster_analysis
-------
 
The EvidentialGene tr2aacds program implements a classifier algorithm to place 
coding-sequence transcripts into predetermined categories:
	primary with alternates (= main), 
	primary, no alternates (= noclass), 
	alternates, with high and medium alignment to primary (= althi1, althi, altmid)
	alternates can have fragment qualifier, shorter than primary (= altfrag)
	alternates can also have protein alignment qualifier (=a2, aa-high-ident), not quite same as CDS alignment.
	partial, short/incomplete alternate transcripts (= part)

Additionally, an okay (keep) or drop value is assigned using scores of alignment and
protein quality (size, completeness), and optionally protein homology, to partition 
useful and not useful transcripts.

In this case the classifier is deterministic, based on absolute transcript qualities, 
so that same results are obtained with different groupings of input data 
(e.g using several rounds to classify subsets, then classifying those 
subset results into a broader group, versus classifying all transcripts from
all treatments into one broad group at one step).  Other kinds of classifiers
produce different outcomes based on how you do your steps, are dependent on
input data groups.  Clustering is also depending on your input data organization.

The drop class contains redundant and uninformative mRNA transcripts versus okay class, 
including (a) perfect duplicates of okay transcripts, (b) perfect fragments of okay transcripts,
and (c) other above classes where additional transcript qualities indicate
an uninformative coding sequence (i.e. protein identity with minor cds difference,
protein/cds fragment with minor difference from okay transcripts).  

The tr2aacds output table myseq.trclass contains these classings with values, per input transcript.
EvidentialGene program evgmrna2tsa2.pl produces a simplified classification table of 
all input transcripts as to above groups from myseq.trclass, as publicset/myseq.mainalt.tab,
along with other publication sequence files.

Parameters in tr2aacds that assign class have been tested and adjusted with known transcript
sets to match those, which maximizes true positive classification.  Basic
parameters for this include sequence length (e.g. minimum of 30 aa, exon local
alignment identity of >= 98%, to distinguish identical shared exon spans from
paralogous or population locus variants).

See also 
http://arthropods.eugenes.org/EvidentialGene/about/EvidentialGene_trassembly_pipe.html
As described here, tr2aacds uses these steps to efficiently classify large input sets
of mRNA transcripts.
  1. perfect redundant removal: exonerate/fastanrdb input.cds > input_nr.cds
  .. uses added aa.qual info for choosing best cds among identicals.
  2. perfect fragment removal: cd-hit-est -c 1.0 -l MINCDS ..
  3. blastn, basic local align hi-ident subsequences for alternate tr., with -perc_identity CDSBLAST_IDENT (98%)
    to find high-identity exon-sized alignments. 
  4. classify main/alternate cds, okay & drop subsets, using evigene/rnaseq/asmrna_dupfilter2.pl
  .. merges alignment table, protein-quality and identity, to score okay-main, ok-alt, and drop sets.

-----