(a) The cases and controls used for training and validation were
drawn from different European population distributions, in a manner consistent
with the differences in minor allele frequencies of the top SNPs in the
classifier. Plotted on rotated and inverted geographic axes of European
variation (Yang et al. 2012) are
individuals from the control population CEU (light blue diamonds representing
Northern and Western European-Americans), and individuals with and without
autism from the training and replication case datasets (AGRE probands as red
asterisks and unaffected relatives as green Xs; SSC probands as red triangles
and unaffected relatives as green circles). The eleven HapMap3 populations and a
reference Northeastern European population (Finns from the 1000 Genomes Project)
are denoted by diamonds. To reduce population stratification bias, Skafidas and
colleagues only used individuals nearer CEU than any other HapMap population to
train the classifier. However, we highlight considerable differences in
population structure between cases and controls even in this subset (represented
in the inset box, the red cross indicates the centroid of CEU). (b)
The weights of the top reported SNPs (x-axis) are correlated
with the minor allele frequency in Estonians (a Northeastern European
population) divided by that of CEU (y-axis)
(p= 4×10-4). While 1000 Genomes data
were not available for Estonians, the Estonian minor allele frequencies are
based on considerably more chromosomes, allowing for more accuracy.
(c) In contrast, the weights of the top reported SNPs
(x-axis) are uncorrelated with the odds ratios among an
independent set of individuals in AGRE nearest CEU (p =
0.997). (d) The case sets are more diverse than the control sets
(p=6.2×10-5 for CEU vs AGRE or
p=4.6×10-8 for CEU vs SSC, two-sided
Kolmogorov-Smirnov test). Cumulative proportion of AGRE, SSC and CEU individuals
plotted against Euclidean distance from the CEU centroid, for those individuals
nearer the CEU centroid than that of any other HapMap population in the inset of
panel a. (e) Frequency distributions of autism risk
scores for the independent set of AGRE cases and controls and four populations
sequenced in the 1000 Genomes Project (FIN – Finns, GBR –
British, TSI – Italians, and CEU – Northwestern Europeans).All
classifier scores were calculated using the 19 SNPs that had data in all of
these sets and were also in the top 30 most discriminative SNPs in Skafidas et al (2012). Neither the
distributions of cases and controls, nor the proportion above the threshold
level differ significantly from one another. In contrast, different populations
have different distributions. Finns differ from neither independent AGRE cases
nor independent AGRE controls, while other European populations have lower
scores.