Ejemplo n.º 1
0
#                   |    |    |          /---|          /---|     \-Hsa0010733
#                   |     \---|         |    |         |    |
#                   |         |         |    |     /---|     \-Hsa0010710
#                   |         |         |    |    |    |
#                   |         |     /---|     \---|     \-Ptr0000001
#               /---|         |    |    |         |
#              |    |         |    |    |          \-Cfa0016699
#              |    |         |    |    |
#              |    |          \---|    |     /-Rno0030248
#              |    |              |     \---|
#              |    |              |          \-Mms0024821
#          /---|    |              |
#         |    |    |               \-Mdo0014718
#         |    |    |
#         |    |    |     /-Dre0008389
#     /---|    |     \---|
#    |    |    |          \-Fru0004507
#    |    |    |
#    |    |     \-Cin0011238
#----|    |
#    |     \-Aga0007658
#    |
#    |--Dme0014628
#    |
#     \-Ddi0002240
#
# Gets the best evolutionary model tree present in the phylome 1 for a given geneid
winner_model, all_likelihoods, t = p.get_best_tree(geneid, 1)
# As you can see, the best likelihood tree in this case was
# reconstructed using a JTT model rather than the WAG matrix.
Ejemplo n.º 2
0
#                   |    |    |          /---|          /---|     \-Hsa0010733
#                   |     \---|         |    |         |    |
#                   |         |         |    |     /---|     \-Hsa0010710
#                   |         |         |    |    |    |
#                   |         |     /---|     \---|     \-Ptr0000001
#               /---|         |    |    |         |
#              |    |         |    |    |          \-Cfa0016699
#              |    |         |    |    |
#              |    |          \---|    |     /-Rno0030248
#              |    |              |     \---|
#              |    |              |          \-Mms0024821
#          /---|    |              |
#         |    |    |               \-Mdo0014718
#         |    |    |
#         |    |    |     /-Dre0008389
#     /---|    |     \---|
#    |    |    |          \-Fru0004507
#    |    |    |
#    |    |     \-Cin0011238
#----|    |
#    |     \-Aga0007658
#    |
#    |--Dme0014628
#    |
#     \-Ddi0002240
#
# Gets the best evolutionary model tree present in the phylome 1 for a given geneid
winner_model, all_likelihoods, t  = p.get_best_tree(geneid, 1)
# As you can see, the best likelihood tree in this case was
# reconstructed using a JTT model rather than the WAG matrix.
Ejemplo n.º 3
0
from ete3 import PhyloTree, PhylomeDBConnector, SeqGroup

p = PhylomeDBConnector()
w,x, t =  p.get_best_tree("Hsa0000001", 1)
a, l = p.get_clean_alg("Hsa0000001", 1)
A = SeqGroup(a, "iphylip")
for s in A.id2seq:
    A.id2seq[s]=A.id2seq[s][:30]
t.link_to_alignment(A)
print t.get_species()
print t
t.set_outgroup(t&"Ddi0002240")

sp = PhyloTree("(((((((((((Hsa, Ptr), Mmu), ((Mms, Rno), (Bta, Cfa))), Mdo), Gga), Xtr), (Dre, Fru))),Cin) (Dme, Aga)), Ddi);")
reconciled, evs = t.reconcile(sp)
print reconciled
reconciled.show()
# This connects to the main phylomeDB server (default parameters)
p = PhylomeDBConnector()
PHYLOME_ID = 1
# This is the species code/age dictionary used to correctly root the
# tree in the human phylome. You can define your own, or use the
# midpoint outgroup method
species2age = {'Aga': 8, 'Ago': 9, 'Ame': 8, 'Ath': 10, 'Bta': 3, 'Cal': 9, 'Cbr': 8,\
             'Cel': 8, 'Cfa': 3, 'Cgl': 9, 'Cin': 7, 'Cne': 9, 'Cre': 10, 'Ddi': 10, \
             'Dha': 9, 'Dme': 8, 'Dre': 6, 'Ecu': 9, 'Fru': 6, 'Gga': 4, 'Gth': 10,\
             'Gze': 9, 'Hsa': 1, 'Kla': 9, 'Lma': 10, 'Mdo': 3, 'Mms': 3, 'Mmu': 2,\
             'Ncr': 9, 'Pfa': 10, 'Pte': 10, 'Ptr': 2, 'Pyo': 10, 'Rno': 3, 'Sce': 9,\
             'Spb': 9, 'Tni': 6, 'Xtr': 5, 'Yli': 9      }
# Iterator over each sequence in the human proteme
for i, seqid in enumerate(p.get_seed_ids(PHYLOME_ID)):
    if i>2: break # Just process the first 2 ids
    winner_model, lks, t = p.get_best_tree(seqid, PHYLOME_ID)
    # If tree was sucsesfully reconstructed, runs the species overalp algorithm
    if t and seqid in t:
        outgroup = t.get_farthest_oldest_leaf(species2age)
        # Returned outgroup is used to root the tree
        t.set_outgroup(outgroup)
        # Finds the node representing the seed sequence.
        # We want the orthology relationships of such sequence.
        seed_node = t.search_nodes(name=seqid)[0]
        evol_events = seed_node.get_my_evol_events()
        for ev in evol_events:
            # Speciation event
            if ev.etype == "S":
                inparalogs = filter(lambda n: n.startswith("Hsa"), ev.in_seqs)
                print 'ORTHOLOGY RELATIONSHIP:', ','.join(inparalogs), "<===>", ','.join(ev.out_seqs)
from ete3 import PhyloTree, PhylomeDBConnector, SeqGroup

p = PhylomeDBConnector()
w, x, t = p.get_best_tree("Hsa0000001", 1)
a, l = p.get_clean_alg("Hsa0000001", 1)
A = SeqGroup(a, "iphylip")
for s in A.id2seq:
    A.id2seq[s] = A.id2seq[s][:30]
t.link_to_alignment(A)
print t.get_species()
print t
t.set_outgroup(t & "Ddi0002240")

sp = PhyloTree(
    "(((((((((((Hsa, Ptr), Mmu), ((Mms, Rno), (Bta, Cfa))), Mdo), Gga), Xtr), (Dre, Fru))),Cin) (Dme, Aga)), Ddi);"
)
reconciled, evs = t.reconcile(sp)
print reconciled
reconciled.show()