# $Id: Humangenefinding.txt,v 1.4 1995/07/17 16:06:37 colin Exp colin $ 

Notes on finding H. sapiens Genes.

A number of parameters should be set on genefinder, since the default
values apply to C. elegans, and appropriate values for H. sapiens are
quite different.  These parameters apply to the intron lengths, relative
weighting, and cutoff values.

-maxintronlength 10500	    (Human introns can be quite long, longer values ok)
-minintronlength 70	    (Minimum intron length in humans ~70 nucleotides)
-penaltycluster 50	    (Intron lengths seem to peak at ~120 =70+50 )

If you are using amino acid or codon tables, genefinder does not work
very well.  There does not appear to be enough information to
distinguish coding sequence very well.  (You can use the information
program to compute the expected score from a single codon in sequence
with the tables you choose to use)

If you deccide to use amino acid tables anyway, I would suggest
changing the intron penalty parameters suggested below to 
-penaltyfactor 1.0 -cpenalty 4.6  
(This reduces the penalty for an intron somewhat to scale with the
reduced score given to coding regions)

Experimentally we determined that the optimal values for penalty
parameters for intron lengths were different for codon tables, amino
acid tables or n-mer tables.  The best results with n-mer tables was
produced with values of:

-penaltyfactor 0.85
-cpenalty 6.35

Unfortunately, genes which have a short exons separated by short
introns will not be predicted by genefinder with these parameters, due
to the large constant penalty (cpenalty) assigned to each intron.

If a large tract of genomic sequence is being analyzed (>10kb) then
better predictions will be made using the -norm flag to compute the
likelihood relative to the genomic sequence instead of random sequence.
As the genomic sequence contains some coding region, a correction factor
is used to adjust the likelihood.  The suggested flags are:

-norm -corrfactor 1.325

If hexamer tables are used, very large regions of sequence are required to
estimate the null values.  In this case it may be appropriate to use the
random sequence.  For this case a correction factor of 0.8 actually works
better in many cases than the default of 1.0, by reducing the false positive
rate.

It may be desirable to replace repeat regions with 'X' to prevent
predictions from occuring there.  With our test set of genes there was a
small overall improvement in gene prediction after repeat regions were
eliminated from the input sequence.

Recommendations:  
For very large sequence (>100kb) analysis, use hexamer tables

genefinder -norm -corrfactor 1.325 -tablenamefile humtables.hex
	-cpenalty 6.35 -penaltyfactor 0.85 -maxintronlength 10500
	-minintronlength 70 -penaltycluster 50


For smaller sequences (10kb-100kb), use pentamer tables

genefinder -norm -corrfactor 1.325 -tablenamefile humtables.pent
	-cpenalty 6.35 -penaltyfactor 0.85 -maxintronlength 10500
	-minintronlength 70 -penaltycluster 50


For small sequence (<10kb) use hexamer tables without normalizing

genefinder -corrfactor 0.8 -tablenamefile humtables.hex
	-cpenalty 6.35 -penaltyfactor 0.85 -maxintronlength 10500
	-minintronlength 70 -penaltycluster 50