targets: (i) those for which evidence of hap-
loinsufficiency as disease mechanism was
concordant with the down-regulation of the
same genes by MIX N and which can be
therefore prioritized for further investigations
toconfirmthemascausativerelaysofEDC
neurodevelopmental outcomes; and (ii) those
for which the direction of dysregulation was
different either across doses (1X and 1000X)
or across models. Given the increasing num-
ber of loci that can cause NDD by both loss or
gain of function (through either copy number
variations or mutations) ( 94 ), we suggest a pre-
cautionary approach for this class of targets
and consider both up- and down-regulation as
potentially adverse, depending on specific times
and concentrations of exposure.
In vivo, MIX N altered TH activity in
Xenopusheads, where the observed reduction
in TH activity acquires human relevance when
placed in the context of epidemiological re-
sults, because slight changes in maternal TH
concentrations during early pregnancy result
in IQ loss and modified brain structure in
offspring ( 95 ). The TH disruptive activity of
MIX N was corroborated in zebrafish, where
we observed a decrease in the expression of
thyroid receptorsaandb. A master regulator
and Genomatix analysis of DEGs revealed
TH and estrogen signaling as the main endo-
crine axes affected by MIX N, in line with the
close link between estrogen and TH signal-
ing in the brain, where many genes are co-
regulated by both hormones ( 96 , 97 ). Finally,
MIX N affected locomotion in both animal
models, providing direct evidence for the im-
pact of EDC mixtures on organismal behavior,
which is of long-established value in toxicol-
ogy. Interestingly, the differential impact on
the two organisms echoes the differences in
species-specific outcomes that was also previ-
ously observed with single chemicals, highlight-
ing how the diversity of species-specific behaviors
(exemplified, for zebrafish andXenopus, in
their differential light-dark cycle locomotion)
may affect the manner in which they respond
to toxic exposures, with relevant implica-
tions, in terms of probing mechanisms, for
both specific chemicals and mixtures in mul-
tiple models ( 98 ).
Our work translates the framework of regu-
latory methods for risk assessment—namely,
combining evidence of human exposure ranges
with experimental evidence of links to adverse
health effects—from single chemicals to com-
plex mixtures. Based on our results, we can
thus recommend (as summarized in Fig. 6) a
risk assessment pipeline attuned to the real-life
impact of chemicals through mixture-centered
biostatistics and experimental biology, along-
side other complementary approaches to mix-
ture assessment. For any given outcome of
concern, as long as reliable mixture-effect
estimates are available alongside relevant
experimental models, we anticipate our ap-
proach to be broadly applicable across health
domains. To maximize feasibility, given its
greater costs over standard single chemical–
based approaches, we suggest a three-pronged
approach entailing (i) a broad benchmarking
effort through which public and private stake-
holders should prioritize together, for each
health domain, the adverse outcomes for
which mixture-based approaches are likely to
enable the greatest strides in public health
protection; (ii) a focused effort on these selected
outcomes and mixtures in diverse populations
through interdisciplinary research consortia;
and (iii) a parallel, finer-grained community
effort, in which the same interdisciplinary
approach can be translated to smaller, more
local settings where experimental and epide-
miological teams are also systematically in-
tegrated from the outset on simpler mixtures
and phenotypes.Materials and methods summary
Exposure assessment
Using data from the SELMA pregnancy cohort
( 13 ), mixtures of prenatal EDC exposures of
relevance for health outcomes in children
were identified. Exposure was measured in
urine and serum taken in weeks 3 to 27 of
pregnancy (median week 10, and 96% of the
samples were taken before week 13). First
morning void urine samples were analyzed for
10 phthalate metabolites [mono-ethyl phthal-
ate (MEP), metabolite of DEP; mono-n-butyl
phthalate (MnBP), metabolite of DBP; mono-
benzyl phthalate (MBzP), metabolite of BBzP;
mono-(2-ethylhexyl) phthalate (MEHP), mono-
(2-ethyl-5-hydroxylhexyl) phthalate (MEHHP),
mono-(2-ethyl-5-oxohexyl) phthalate (MEOHP),
and mono-(2-ethyl-5-carboxypentyl) phthalate
(MECPP), metabolites of DEHP; mono-hydroxy-
iso-nonyl phthalate (MHiNP), mono-, oxo-iso-
nonyl phthalate (MOiNP), and mono-carboxy-
iso-octyl phthalate (MCiOP), metabolites of
DiNP)]; and alkyl phenols, including BPA and
triclosan. Serum samples were analyzed for
eight perfluorinated alkyl acids [perfluorohep-
tanoic acid (PFHpA), perfluorooctanoic acid
(PFOA), perfluorononanoic acid (PFNA), per-
fluorodecanoic acid (PFDA), perfluoroundeca-
noic acid (PFUnDA), perfluorododecanoic acid
(PFDoDA), perfluorohexane sulfonate (PFHxS),
and perfluorooctane sulfonate (PFOS)] as des-
cribed in the supplementary materials and
publications therein.Health examinations
For a measure of neurodevelopment, we used
data from a routinely made language screen-
ing of the children when they were 30 months
old. Language development was assessed by
nurse’s evaluation and parental question-
naire, including the number of words the
child used (<25, 25 to 50, and >50). The mainstudy outcome was the use of fewer than 50
words, termed language delay corresponding
to a prevalence of 10%.Biostatistical analyses
WQS regression ( 16 ) adjusted for covariates
was used to establish associations between
mixture exposures and language delay in
children (see statistical method in the supple-
mentary materials). WQS regression is a strat-
egy for estimating empirical weights for a
weighted sum of quantiled concentrations
(i.e., deciles) most associated with the health
outcome. The results are abcoefficient asso-
ciated with the weighted sum (estimate, stan-
dard error, andPvalue) and the empirical
weights (which are the average weights, con-
strained to sum to 1, from an ensemble step—
here, 100 bootstrap samples). The components
most associated with the health outcomes
have non-negligible weights and were treated
as sEDCs when the estimated weight exceeds
the equi-weight case.
Next, we estimated the equivalent daily in-
take of sEDCs measured in the urine (phtha-
lates and alkyl phenols) and estimated serum
concentrations from the daily intake for these
urinary measurement-based compounds. Finally,
we used the geometric means, on a molar
basis, for either the measured (PFAS) or
estimated serum concentrations (phthalates
and alkyl phenols) and established mixing
proportions to prepare the mixture associated
with language delay (MIX N). This mixture
was used in the experimental studies.Composition of the mixtures
The chemicals needed for the mixture were
obtained from commercial custom synthesis
laboratories or vendors. BPA, DMSO, MBzP,
PFHxS, PFNA, and PFOS were obtained from
Sigma-AldrichInc.MEPandMiNPwereob-
tained from Toronto Research Chemicals. MBP
was purchased from TCI, Tokyo Chemical
Industry Co., Ltd. For MIX N, 1 M solutions
in DMSO were prepared using MEP, MBP,
MBzP, MiNP, BPA, PFHxS, PFNA, and PFOS.
Thereafter, the 1 M solutions were mixed in
proportions as described in table S3.Human cellular models
Human iPSCs have been previously validated
in G.T.’s laboratory ( 99 ). HFPNSCs were de-
rived and maintained as previously described
( 100 ). HFPNSCs were obtained after elective
termination with no evidence of develop-
mental abnormalities. For the differentia-
tion of cortical brain organoids, we used the
protocol set up by ( 101 ), with minor modifica-
tions to improve its efficiency as shown pre-
viouslybyusin( 27 ). Detailed information on
culture condition, experimental procedures, and
quality control are available in the supplemen-
tary materials.Caporaleet al.,Science 375 , eabe8244 (2022) 18 February 2022 11 of 15
RESEARCH | RESEARCH ARTICLE