Monday, March 30, 2015

Inconsequential splits in NeighborNet graphs

NeighborNet produces splits graphs based on distances between the taxa, rather than using the original character data. This approach can produce what we might call inconsequential splits in the graph — that is, splits that are not explicitly supported by the character data. Here, I present a simple example to illustrate the extent to which this can occur.

The data are taken from: Nanette Thomas, Jeremy J. Bruhl, Andrew Ford, Peter H. Weston (2014) Molecular dating of Winteraceae reveals a complex biogeographical history involving both ancient Gondwanan vicariance and long-distance dispersal. Journal of Biogeography 41: 894-904.

This dataset consists of a set of eight morphological features of the pollen from 31 extant plant taxa plus two fossil samples, as shown in this data matrix:

T_lanceolata        00111011
T_stipitata         00111011
T_purpurescens      00111011
T_xerophila_x       00111011
T_xerophila_r       00111011
T_vickeriana        00111011
T_glaucifolia       00111011
T_membranea         00111011
T_insipida          00111011
T_perrieri          00111010
D_winteri           00111010
D_grenadensis       00111010
B_comptonii         00011010
B_howeana           00011010
B_semicarpoides     00011010
B_whiteana          00011010
B_queenslandiana_q  00011010
B_queenslandiana_1  00011010
P_axillaris         00011011
P_colorata          00011011
Pseudowinterapollis 00011011
B_pancheri          01001011
Harrisipollenites   01001100
Z_acsmithii         01001101
E_stipitatum        01001101
Z_bicolor           01001101
Z_balansae          11001101
C_dinisii           1-111101
C_madagascariensis  1-111101
W_salutaris         1-111101
P_macranthum        1-111101
C_ekmanii           1-111101
C_winterana         1-111101

Note that there are only nine groups of taxa (separated by the dashed lines) — within each group the data are identical. Each character has two states: present / absent.

The resulting NeighborNet, as produced by default using the SplitsTree4 program, is shown in the first graph.

As expected, the taxa form nine groups. There are a number of apparently well-supported splits (ie. with long edges) separating these groups. There are also a number of smaller splits, and a whole series of very tiny splits. None of these latter two groupings are explicitly present in the dataset — the only splits supported by the characters are plotted onto the graph using the character numbers. (Note that character 5 is uninformative.)

The series of very tiny splits are present throughout the graph as extremely short edges. For example, a detailed view of the bottom left-hand corner of the graph is shown in the next figure.

Note that these six taxa have identical character data, and therefore their separation into four groups is entirely an artifact of the NeighborNet algorithm.

So, one needs to be careful when interpreting small splits in such a graph — they may have biologiocal support and they may not.

No comments:

Post a Comment