Combining High-Throughput Screening and In Silico Modeling to Derisk Novel Agrochemicals for Androgen Receptor Binding

Androgen receptor (AR) modulation is a critical safety
concern
for environmental chemicals, including agrochemicals, due to its role
in endocrine disruption. Existing public data sets for AR modulation
are limited in size and diversity. Here, we report a large-scale high-throughput
screening (HTS) campaign assessing AR binding for over 72,000 compounds
from an agrochemical library using a fluorescence polarization displacement
assay. Confirmatory dose–response testing identified 4,183
AR binders (5.7% hit rate) with substantial structural diversity and
numerous novel scaffolds. To enable predictive modeling, we curated
an unrestricted data set of 24,953 compounds with associated activity
data, which we publish as part of this paper. Using this data set,
we trained machine learning models based on molecular 1D and 2D descriptors
and fingerprints. Gradient-boosted trees achieved the best performance,
with a balanced accuracy of 0.77 and a negative predictive value of
0.98, making the model suitable for derisking large virtual libraries.
External validation on existing publicly available AR data sets (CoMPARA
and PubChem) demonstrated reasonable transferability (balanced accuracy
0.66 and 0.72), overcoming differences in experimental methods and
composition of the compound sets. Our findings demonstrate the utility
of combining HTS with machine learning for early safety assessment
and provide a benchmark data set to advance AR binding prediction,
complementing existing data sets