3.3 | Random-Forest ABC scenario-choice for the history
of ACB and ASW populations
We performed RF-ABC model-choice separately for the admixture history of
the Barbadian (ACB) and the African American (ASW) populations, to
evaluate whether our MetHis -ABC method could identify subtle
differences in the history of both populations having experienced the
TAST under the British colonial empire (Martin et al. 2017; Baharian et
al. 2016). For the ACB, Figure 3 shows that the majority of
votes (53.1%) went to an admixture scenario AfrDE-EurDE with a
posterior probability of the winning scenario of 60.28%. This posterior
probability is above the mean posterior-probability obtained when the
wrong scenario is chosen for the 1000 AfrDE-EurDE simulations closest to
the observed one (56.8%, SD=11.6%, for 37 simulations wrongly assigned
in total). The second most chosen scenario was the AfrDE-Eur2P scenario.
However, this scenario is voted for 3.5 times less often than the
winning scenario AfrDE-EurDE, gathering 15.1% of the 1,000 votes, only
slightly above the 11.11% prior probability for each nine-competing
scenario (Figure 3 ; Supplementary Table S1 ).
RF-ABC scenario-choice results were less segregating for the ASW.Figure 3 shows that the AfrDE-EurDE scenario also gathered the
majority of votes, albeit with lower posterior probability than for the
ACB (33.5% of 1,000 votes, with posterior probability = 48.0%). This
posterior-probability is slightly below the average
posterior-probability obtained when the wrong scenario is chosen for the
1000 AfrDE-EurDE simulations closest to the ASW (50.7%, SD = 7.9%, for
192 simulations wrongly assigned). The second most chosen scenario,
AfrDE-Eur2P, was only slightly less chosen with 31.7% of the votes
(Figure 3 , Supplementary Table S1 ). Altogether these
results denote an ambiguity of the RF-model choice in the part of the
parameter-space occupied by the ASW.
Considering only these two best scenarios to train the RF and
re-conducting scenario-choice improved the scenario discrimination in
favor of the AfrDE-EurDE scenario. While we found, again, only a slight
majority of votes (51.8%) in favor of the AfrDE-EurDE scenario, the
posterior probability for this model was substantially increased to
57.9%, thus above the average posterior-probability threshold
calculated above (50.7%). This indicated that the AfrDE-EurDE scenario
best explained the ASW observed genetic patterns, despite overall
limited discriminatory power of our approach in the ambiguous part of
the summary-statistics space occupied by this population.