Check CART classification model for nifH about metatag

[Robaina]

CART model to classify nifH

The CART model is taken from this publication.

Departing from the nifH sequence of Azotobacter, it classifies peptides based on the position of three residues following:

    CART = {
        109: ['F', 'W', 'Y'], 
        49: ['A', 'D', 'I'],  
        53: ['L', 'M', 'W'] 
    }

and the decision tree:

Now, we have included Azotobacter's sequence in our nifH peptide database, followed by a MSA of all the peptide sequences in there. To use the CART model, we first have to find the correspondence between the residue positions in the original, unaligned Azotobacter sequence and the aligned one. This is currently done manually since this correspondence changes depending on the set of sequences that the database contains (it changes the MSA). For our nifH database, the correspondence is (unaligned: aligned):

adjusted_CART = {
    109: 180,
    49: 95,
    53: 99
}

Here are the unaligned and aligned nifH sequences of Azotobacter:

unaligned:

>001_WP_039801084.1 MULTISPECIES: nitrogenase iron protein [Azotobacter]
MALRQCAIYGKGGIGKSTTTQNLVAALAEAGKKVMIVGCDPKADSTRLILHSKAQNTVMEMAASAGSVEDLELEDVLQIGYGGVKCVESGGPEPGVGCAGRGVITAINFLEEEGAYSDDLDFVFYDVLGDVVCGGFAMPIRENKAQEIYIVCSGEMMAMYAANNIAKGIVKYAHSGSVRLGGLICNSRKTDREDELIMALAAKIGTQMIHFVPRDNVVQHAEIRRMTVIEYDPKAKQADEYRALAQKILNNKLLVIPNPASMEDLEELLMEFGIMEAEDESIVGKAAAEG

aligned:

>001_WP_039801084.1 MULTISPECIES: nitrogenase iron protein [Azotobacter]_cluster_I 001_WP_039801084.1 MULTISPECIES: nitrogenase iron protein [Azotobacter]_cluster_I
---------------------------------------------MALRQCAIYGKGGIGKSTTTQNLVAALAE-AGKKVMIVGCDPKADSTRLILHSKAQN-TVMEMAASAGSV---------ED-LEL-ED-VLQIG------------YGGVKCVESGGPEPGVGCAGRGVITAINFLEEEG-AYSDD-----LDFVFYDVLGDVVCGGFAMPIRENKAQEIYIVCSGEMMAMYAANNIAKGIVKYAHS--GSVRLGGLICNSRKTDR-EDELIMALAAKIGTQMIHFVPRDNVVQHAEIRRMT------VIEY-------------DPKAKQADEYRALAQK----ILNN--K-L--LVIPNP-ASMED------------LEE------------------------LL--------ME-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------FGIMEA--EDESIV--GKAAAEG--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

The script that implements the CART model can be found in:

https://github.com/Robaina/TRAITS/blob/main/code/classifyNifHsequences.py

NOTE that residue positions in the script are one unit behind because Python follows a 0-indexing scheme.

Using the script above, we obtain the following cluster distribution for the nifH peptide database:

[Robaina]

I have checked both the model and its implementation and they seem fine. Therefore, the observation that too few environmental sequences are classified outside of cluster IV (hence classified as "true" nifH) could be explained by:

1. The CART model itself is incomplete / doesn't capture all possible variability of nifH sequences
2. There CART model is right and there is really a small number of non - cluster IV environmental sequences in some samples
3. A combination of the previous ones...

[Robaina]

Alright, closing this issue because the CART model will no longer be used to cluster nifH