Application of Deep Learning to Predict Human Oral Bioavailability of Pharmaceuticals
Abstract
High
failure
rates
in
drug
development
are
predominantly
driven
by
suboptimal
ADMET
(absorption,
distribution,
metabolism,
excretion,
and
toxicity)
properties,
with
human
oral
bioavailability
(HOB)
serving
as
a
critical
determinant
of
therapeutic
efficacy
safety.
Traditional
HOB
assessment
methods,
reliant
on
animal
models
clinical
trials,
face
inherent
limitations
cost,
scalability,
reproducibility.
To
address
these
challenges,
this
study
proposes
deep
learning
framework
integrating
the
directed
message-passing
neural
network
(D-MPNN)
from
Chemprop
tool
RDKit-derived
molecular
descriptors,
enhancing
predictive
accuracy
through
hybrid
representations
atomic/bond-level
graph
features
global
physicochemical
properties.
Bayesian
optimization
automated
hyperparameter
tuning,
while
ensemble
(20
models)
ensured
robustness
for
model
development.
The
optimized
achieved
an
AUC
0.8299
77.65%
internal
validation,
outperforming
existing
tools
75
%
external
FDA-approved
drugs.
Interpretability
analysis
identified
substructures
correlated
high
HOB,
providing
actionable
insights
rational
design.
This
work
establishes
novel
method
high-throughput
screening
candidates
favorable
bioavailability,
highlighting
potential
to
decode
complex
structure-property
relationships
pharmaceutical
optimization.
Published: May 21, 2025
Language: Английский