exposure and language delay in children, on
the basis of the original hypothesis on the di-
rection of the outcome associated with pre-
natal EDC mixture exposure (delay rather than
enhancement of language development). The
focus of WQS regression is not on individual
components where hypothesis testing and ef-
fect sizes are generally of interest. Instead, its
aim is to detect components of interest (sEDCs)
through non-negligible weights, particularly
when the components have complex correla-
tion patterns. The effect size of the mixture is
measured in terms of a mixture effect of a
weighted index of deciled components. So, a
unit increase in the weighted index has an
effect size ofb. In a logistic model with a
binary outcome such as language delay, the
effect size is exp(b) as an odds ratio. For
language delay, the odds ratio was roughly
1.4 across multiple WQS analyses. WQS regres-
sion was then adjusted to control for the
following relevant potential confounders (though
there may conceivably be residual confounding
fromvariablesthatwewereunabletocollect):
sex of the child, smoking status for the mother,
parity, fish consumption, maternal educa-
tion, and creatinine concentrations (for uri-
nary metabolite concentrations). Adjustment
by these potential confounders increased the
odds ratio from 1.2 to 1.4. As in all regression
models, WQS regression suffers from the im-
pact of unmeasured confounders.
For the experimental identification of the
genes dysregulated by MIX N exposure, our
complimentary analysis in fetal progenitors
and brain organoids demonstrated the feasi-
bility of translating epidemiological evidence
into mechanistic readouts in pathophysiolog-
ically relevant human models. This allowed us
to draw the following conclusions and chal-
lenges ahead.
First, whereas the two models shared con-
cordant clusters of MIX N dysregulated genes,
they also captured different and, at times,
opposing aspects of MIX NÐinduced dysre-
gulation, underscoring the importance of
endophenotype-based analysis across human
models and longitudinal alignment between
high-resolution clinical and molecular data
( 93 ). Thus, whereas the enrichment of catego-
ries related to protein translation was mainly
driven by genes that were dysregulated only in
fetal progenitors, organoids uncovered dysre-
gulation of specific neuronal functions such
as axonogenesis and synaptic control (fig. S8).
This functional partitioning is consistent withthe inherent difference between models (self-
renewal of progenitors in fetal progenitors versus
actual development in organoids) and indi-
cates that MIX N alters neurodevelopment by
affecting both translational control in progen-
itors and more mature aspects of neuronal cell
biology at later stages.
Second, organoids were more sensitive to
EDC exposure, showing more DEGs upon MIX
N exposure and more significant (P< 0.05)
overlaps with NDD causative genes. Together
with the above observation on stage-specific
dysregulation in organoids versus fetal pro-
genitors, this underscores the complementa-
rity of the two systems in capturing different
temporal windows of neurodevelopmental
vulnerability. Thus, although fetal progenitors
may have been considered a gold standard for
developmental biology and toxicology because
they are directly derived from primary tissue,
our results show that brain organoids are not
simplyamoreviablealternative(becauseof
the inherent procurement issues associated
with fetal material) but also a more sensitive
and pathophysiologically relevant model.
Third, the integration of MIX N DEGs with
comprehensive lists of NDD-causative genes al-
lowed us to pinpoint two classes of molecularCaporaleet al.,Science 375 , eabe8244 (2022) 18 February 2022 10 of 15
EPIDEMIOLOGYMIX NBIOSTATISTICSCHEMISTRYEXPERIMENTALBIOLOGYPhthalatesPFASbisphenol-A0.25 0.20 0.15 0.10 0.05 0SMACHFetal progenitors Brain organoidsWQSSelma Cohort Motivation: Epidemiological data and human biomonitoring should guide
experimental toxicology to include the concentrations of chemical mixtures
relevant for real-life exposure and for which there is evidence of associations
between exposures and health outcomes of concern.Step 1: Biostatistical methods for characterizing environmental mixtures such
as weighted quantile sum (WQS) regression should be used to identify
combinations of chemicals that are associated with health outcomes in humans.
The selection of chemicals should be both sensitive (identifying chemicals of
concern) and specific (identifying chemicals that are not of concern).Step 2: One or more human-relevant typical mixtures (relative proportions and
total concentrations) should be identified, synthesized and experimentally
tested.Step 3: Experimental evidence should identify the molecular mechanisms of
action of the mixtures associated with adverse outcomes and dose-response
experiments should be performed for estimation of a Point of Departure (POD),
through the integration of:human experimental systems, especially induced pluripotent stem cell (iPSC)
derived organoids that recapitulate the tissue-level complexity and
developmental timings of human exposure
in vivo models validated by the OECDStep 4:similar mixture approach (SMACH) should be applied to compare
human exposure (determined to be sufficiently similar to the experimental
mixture(s)) to experimental evidence of the mixture POD using the similar
mixture risk index (SMRI) and determine the proportion of the human population
with exposure ranges of concern (SMRI>1).Conclusion:BasedontheintegrationofevidencefromStep 1 through 4
standard single chemical risk assessment strategies should be benchmarked
against the backdrop of relevant exposure to chemical mixtures of concern.EDCMixRiskFig. 6. Strategic guidelines for chemical mixtures risk assessment.
RESEARCH | RESEARCH ARTICLE