Sample
efficiency
is
a
fundamental
challenge
in
de
novo
molecular
design.
Ideally,
generative
models
should
learn
to
satisfy
desired
objectives
under
minimal
oracle
evaluations
(computational
prediction
or
wet-lab
experiment).
This
problem
becomes
more
apparent
when
using
oracles
that
can
provide
increased
predictive
accuracy
but
impose
significant
cost.
Molecular
have
shown
remarkable
sample
coupled
with
reinforcement
learn-
ing,
as
demonstrated
the
Practical
Optimization
(PMO)
benchmark.
Here,
we
propose
novel
algorithm
called
Augmented
Memory
combines
data
augmentation
experience
replay.
We
show
scores
obtained
from
calls
be
reused
update
model
multiple
times.
compare
previously
proposed
algorithms
and
significantly
enhanced
an
exploitation
task
drug
discovery
case
study
requiring
both
exploration
exploitation.
Our
method
achieves
new
state-of-the-art
PMO
benchmark
which
enforces
computational
budget,
outperforms
previous
best
performing
on
19/23
tasks.
Chemical Reviews,
Год журнала:
2024,
Номер
124(16), С. 9633 - 9732
Опубликована: Авг. 13, 2024
Self-driving
laboratories
(SDLs)
promise
an
accelerated
application
of
the
scientific
method.
Through
automation
experimental
workflows,
along
with
autonomous
planning,
SDLs
hold
potential
to
greatly
accelerate
research
in
chemistry
and
materials
discovery.
This
review
provides
in-depth
analysis
state-of-the-art
SDL
technology,
its
applications
across
various
disciplines,
implications
for
industry.
additionally
overview
enabling
technologies
SDLs,
including
their
hardware,
software,
integration
laboratory
infrastructure.
Most
importantly,
this
explores
diverse
range
domains
where
have
made
significant
contributions,
from
drug
discovery
science
genomics
chemistry.
We
provide
a
comprehensive
existing
real-world
examples
different
levels
automation,
challenges
limitations
associated
each
domain.
JACS Au,
Год журнала:
2025,
Номер
5(5), С. 2294 - 2308
Опубликована: Апрель 23, 2025
Deep
generative
models
yielding
transition
metal
complexes
(TMCs)
remain
scarce
despite
the
key
role
of
these
compounds
in
industrial
catalytic
processes,
anticancer
therapies,
and
energy
transition.
Compared
to
drug
discovery
within
chemical
space
organic
molecules,
TMCs
pose
further
challenges,
including
encoding
bonds
higher
complexity
need
optimize
multiple
properties.
In
this
work,
we
developed
a
model
for
inverse
design
ligands
complexes,
based
on
junction
tree
variational
autoencoder
(JT-VAE).
After
implementing
SMILES-based
metal-ligand
bonds,
was
trained
with
tmQMg-L
ligand
library,
allowing
generation
thousands
novel,
highly
diverse
monodentate
(κ1)
bidentate
(κ2)
ligands,
imines,
phosphines,
carbenes.
Further,
generated
were
labeled
two
target
properties
reflecting
stability
electron
density
associated
homoleptic
iridium
TMCs:
HOMO-LUMO
gap
(ϵ)
charge
center
(q
Ir).
This
data
used
implement
conditional
that
from
prompt,
single-
or
dual-objective
optimizing
either
both
ϵ
q
Ir
interpretation
optimization
trajectories.
The
optimizations
also
had
an
impact
other
properties,
dissociation
energies
oxidative
addition
barriers.
A
similar
implemented
condition
by
solubility
steric
bulk.
Journal of Chemical Information and Modeling,
Год журнала:
2023,
Номер
63(12), С. 3659 - 3668
Опубликована: Июнь 14, 2023
Machine
learning
models
are
increasingly
being
utilized
to
predict
outcomes
of
organic
chemical
reactions.
A
large
amount
reaction
data
is
used
train
these
models,
which
in
stark
contrast
how
expert
chemists
discover
and
develop
new
reactions
by
leveraging
information
from
a
small
number
relevant
transformations.
Transfer
active
two
strategies
that
can
operate
low-data
situations,
may
help
fill
this
gap
promote
the
use
machine
for
tackling
real-world
challenges
synthesis.
This
Perspective
introduces
transfer
connects
potential
opportunities
directions
further
research,
especially
area
prospective
development
Chemical Science,
Год журнала:
2024,
Номер
unknown
Опубликована: Янв. 1, 2024
Evolutionary
and
machine
learning
methods
have
been
successfully
applied
to
the
generation
of
molecules
materials
exhibiting
desired
properties.
The
combination
these
two
paradigms
in
inverse
design
tasks
can
yield
powerful
that
explore
massive
chemical
spaces
more
efficiently,
improving
quality
generated
compounds.
However,
such
synergistic
approaches
are
still
an
incipient
area
research
appear
underexplored
literature.
This
perspective
covers
different
ways
incorporating
into
evolutionary
frameworks,
with
overall
goal
increasing
optimization
efficiency
genetic
algorithms.
In
particular,
surrogate
models
for
faster
fitness
function
evaluation,
discriminator
control
population
diversity
on-the-fly,
based
crossover
operations,
evolution
latent
space
discussed.
further
potential
generative
is
also
assessed,
outlining
promising
directions
future
developments.
PeerJ Physical Chemistry,
Год журнала:
2025,
Номер
7, С. e34 - e34
Опубликована: Янв. 6, 2025
This
study
introduces
a
novel
approach
for
the
de
novo
design
of
transition
metal
catalysts,
leveraging
power
genetic
algorithms
and
density
functional
theory
calculations.
By
focusing
on
Suzuki
reaction,
known
its
significance
in
forming
carbon-carbon
bonds,
we
demonstrate
effectiveness
fragment-based
graph-based
identifying
ligands
palladium-based
catalytic
systems.
Our
research
highlights
capability
these
to
generate
with
desired
thermodynamic
properties,
moving
beyond
restriction
enumerated
chemical
libraries.
Limitations
applicability
machine
learning
models
are
overcome
by
calculating
properties
from
first
principle.
The
inclusion
synthetic
accessibility
scores
further
refines
search,
steering
it
towards
more
practically
feasible
ligands.
Through
examination
both
palladium
alternative
catalysts
like
copper
silver,
our
findings
reveal
algorithms’
ability
uncover
unique
catalyst
structures
within
target
energy
range,
offering
insights
into
electronic
steric
effects
necessary
effective
catalysis.
work
not
only
proves
potential
cost-effective
scalable
discovery
new
but
also
sets
stage
future
exploration
predefined
spaces,
enhancing
toolkit
available
design.
