The first 5 pages, shown as text so search engines and readers can find this document.
Page 1
3808 | Chem. Soc. Rev., 2017, 46, 3808--3829 This journal is © The Royal Society of Chemistry 2017Cite this: Chem. Soc. Rev., 2017,
46, 3808Physical principles for DNA tile self-assembly
Constantine G. Evans a and Erik Winfree *b
DNA tiles provide a promising technique for assembling structures with nanoscale resolution through
self-assembly by basic interactions rather than top-down assembly of individual structures. Tile systems
can be programmed to grow based on logical rules, allowing for a small number of tile types to
assemble large, complex assemblies that can retain nanoscale resolution. Such algorithmic systems can
even assemble different structures using the same tiles, based on inputs that seed the growth. While
programming and theoretical analysis of tile self-assembly often makes use of abstract logical models of
growth, experimentally implemented systems are governed by nanoscale physical processes that can
lead to very different behavior, more accurately modeled by taking into account the thermodynamics
and kinetics of tile attachment and detachment in solution. This review discusses the relationships
between more abstract and more physically realistic tile assembly models. A central concern is how
consideration of model differences enables the design of tile systems that robustly exhibit the desired
abstract behavior in realistic physical models and in experimental implementations. Conversely, we
identify situations where self-assembly in abstract models can not be well-approximated by physically
realistic models, putting constraints on physical relevance of the abstract models. To facilitate the
discussion, we introduce a unified model of tile self-assembly that clarifies the relationships between
several well-studied models in the literature. Throughout, we highlight open questions regarding the
physical principles for DNA tile self-assembly.
1 Introduction
Molecules that self-assemble into macromolecular structures are
found throughout nature. These structures can range from simple
homogeneous crystals to complex biological structures such as
cytoskeletal microtubules of tubulin, virus capsids, and actin
filaments.1–3 The simplicity and power of such self-assembly
is striking: while many synthetic structures, like houses or inte-
grated circuits, are fabricated individually by outside processes,
self-assembling structures are able to form from their basic
components as a consequence of physical principles, often
allowing a multitude of structures to form simultaneously.
The degree of complexity possible in self-assembled structures
is a theoretically compelling question. From a mathematical
perspective, the study of tiling theory made seminal contributions
to the understanding of periodic crystalline structures, such as
the classification of all two-dimensional and three-dimensional
symmetry groups.4 Tiling theory also uncovered more complex, and
often beautiful, possibilities, such as the existence of aperiodic
tilings related to quasicrystals.5,6 Perhaps more surprising, a direct
correspondence between tilings and Turing machines7,8 showed
that the geometry of fitting tiles together can induce patterns of
remarkable ‘‘algorithmic’’ complexity. There are even finite sets of
tiles that can tile the infinite plane (in the sense that mathematical
proof shows that a perfect tiling exists) but for which the resulting
pattern is non-algorithmic (in the sense that no Turing machine or
other computer can calculate how to correctly place the tiles).9,10
This ‘‘unreasonable effectiveness’’ of tiling theory suggests that
considerations of perfect tilings (equivalently, thermodynamic
ground states) do not adequately describe what structures can
and will actually form in a physical system. Consequently,
complexity questions about self-assembly involve not just thermo-
dynamic principles, but must also carefully treat the kinetics.
A theory of tile self-assembly, which treats not the existence of
tilings but rather their growth from a seed according to general-
ized crystal growth principles, has been developed and studied
extensively;11,12 while the ‘‘unreasonable effectiveness’’ of existen-
tial tiling theory is eliminated, the connection to Turing machine
computation and algorithmic patterns remains intact, and the
theory predicts that remarkably complex structures can be self-
assembled even from relatively small sets of tiles.
In theories of generalized crystal growth,16,17 we can identify
(at least) three different classes of self-assembly.
Periodic self-assembly includes classical crystal growth in
which the same element or set of elements are arranged
a Evans Foundation for Molecular Medicine and California Institute of Technology,
Physics, Pasadena, CA, USA
b California Institute of Technology, Computer Science and California Institute of
Technology, Bioengineering, Pasadena, CA, USA. E-mail: winfree@caltech.edu
Received 15th October 2016
DOI: 10.1039/c6cs00745g
rsc.li/chem-soc-rev
Chem Soc Rev
REVIEW ARTICLEPublished on 10 May 2017. Downloaded by California Institute of Technology on 4/24/2026 7:33:00 PM.View Article OnlineView Journal| View Issue
Page 2
This journal is © The Royal Society of Chemistry 2017 Chem. Soc. Rev., 2017, 46, 3808--3829 | 3809periodically to create an object of unbounded size, as well as
analogous examples from biology such as microtubules that self-
assemble from a- and b-tubulin.3,18 Thus, each monomer type
appears within the final structure an unbounded number of times.
Uniquely-addressed self-assembly is in some sense the
opposite extreme, wherein each monomer type appears exactly
once and has binding interactions that each dictate a specific
neighbor. An example from biology would be the eukaryotic
ribosome, which self-assembles from nearly a hundred distinct
RNA and protein components.
Somewhere in between would be algorithmic self-assembly,
which actually encompasses the prior two classes as trivial
cases, but which also includes situations where the some or
all monomer types appear multiple times in the final structure.
Depending on the tile set, algorithmic self-assembly may
produce finite structures (which, unlike uniquely-addressed
self-assembly, may be much larger than the total number of
monomer species) or may produce structures of unbounded
size (which, unlike periodic self-assembly, may create patterns
of arrangement with no periodicity). The self-assembly of
complex virus capsids (such as T4 with its icosahedral head,
whiskers, tail, baseplate, and tail fibers19) could be considered
biological examples of algorithmic self-assembly, in the sense
that each distinct protein monomer species appears multiple
times in the final structure and some kind of logic during self-
assembly is required to determine whether a given monomer
will bind to one potential neighbor species or another. The
quintessential form of logic used in tile-based algorithmic self-
assembly is cooperative attachment, where a certain tile type
binds to a growing assembly only in locations where it can
make a particular pair of binding interactions. This is sufficient
for growth processes that embody Turing machine computa-
tions and other complex algorithmic growth behaviors.
Algorithmic self-assembly can be implemented physically using
DNA tiles, which have been used to create numerous structures of
increasing complexity,20 as illustrated in Fig. 1. DNA tiles, which
can have a variety of structures, attach to each other by short,
single-stranded regions of typically 5 to 10 nucleotides, commonly
referred to as ‘‘sticky ends.’’ The complementary binding of DNA
allows binding rules to be programmed by setting the sticky end
sequences of each tile type, resulting in programmable self-
assembly that can demonstrate many of the theoretically interest-
ing behaviors. The systems can be made to have considerable
complexity: even short sticky ends of 5 nucleotides (like those
found on double-crossover (DX) tiles21,22) can allow 30 to 60
unique sticky end sequences, and longer sticky ends (like those
found on single-stranded tiles (SSTs)23) can allow for considerably
more, according to plausible sequence design criteria.24 DNA tiles
have been used in a wide range of large periodic structures20 and
structures with uniquely-addressed tiles that have ranged up to a
thousand unique tiles in three dimensions.14,25
Experimental implementation of algorithmic self-assembly,
however, has posed several challenges. Of primary importance,
two types of assembly defects can occur that result in undesired
structures: growth errors where tiles attach in locations where
they have imperfect bond matches, and spurious nucleation
where tiles assemble without a seed.26–30 Considerable progress
has been made in overcoming these challenges, and increasingly
complex structures have been successfully built.30,31 This pro-
gress has benefited from refinements in experimental technique,
but derives more significantly from a better understanding of the
physics behind DNA tile assembly: simple models of growth have
motivated fundamental changes to tile system design and
experimental protocols.
In this review, we examine models of DNA tile growth, both
more abstract and more physically-motivated, and show how
Fig. 1 Examples of types of DNA tile self-assembly. (a) Shows the abstract representation, tile structure, lattice representation and atomic force
microscope (AFM) image of a periodic tile system of DAO-E tiles;13 different shades represent end/complement sticky end pairs. (b) Shows a lattice
representation and two AFM image of a uniquely-addressed tile system, where subsets of tiles that assemble into a square have been added in order to
assemble different shapes.14 (c) Shows a representation and AFM image of a large, algorithmic tile system implementing a binary counter ribbon.15 Figure
components adapted from ref. 13–15, with permission of authors.
Review Article Chem Soc RevPublished on 10 May 2017. Downloaded by California Institute of Technology on 4/24/2026 7:33:00 PM.View Article Online
Page 3
3810 | Chem. Soc. Rev., 2017, 46, 3808--3829 This journal is © The Royal Society of Chemistry 2017tile system behavior changes when moving from model to
model. We begin in Section 2 with the example of systems
designed to grow in a specific abstract model, the abstract
Tile Assembly Model (aTAM), which has been well-studied in
theoretical works and experimental implementations. In sub-
sequent sections, we then consider the phenomena that arise
when transitioning through a series of more physically accurate
models, and how the effects of these phenomena can be
mitigated through design and growth constraints in order to
implement experimental systems that behave in approximately
equivalent ways to their abstract counterparts.
In Sections 2 and 3, we show how the transition from the
aTAM to a kinetically-derived model, the kinetic Tile Assembly
Model (kTAM), allows errors in growth to arise, and how these
can be modelled through kinetic and thermodynamic methods
and ameliorated through tile system design methods referred
to as ‘‘proofreading.’’ In Section 4, we consider the nucleation
phenomena that must be addressed when transitioning from
models that start from pre-determined seeds to models that can
start from any tile–tile attachment event, and explain methods
that have been developed to control tile system nucleation.
In Section 5, we explain the challenges to preserving tile system
behavior when transitioning to a model that considers the
depletion of free tile monomers, and explain experimental
methods that have been used to avoid such depletion and
complexities.
We then examine more formally in Section 6 the relation-
ships between models by presenting a unified framework for
models of tile assembly. Choices of reactions and parameters in
this framework can result in models equivalent to or closely
related to several well-studied abstract models, like the aTAM
and the Two-Handed Assembly Model (2HAM), as well as more
physically accurate variations, like the kinetic Tile Assembly
Model (kTAM). Using this framework, we briefly discuss some
of the phenomena that arise when considering tile systems
in variants of the aTAM, 2HAM, and other models, and the
constraints these phenomena place on the tile systems’ physical
relevance.
2 Basic models
Two complementary models of tile assembly are widely studied,
providing a starting point for discussion.26 The abstract Tile
Assembly Model (aTAM) derives from the basic concept of
tile attachment via matching bonds; it provides insight into
the potential computational power of algorithmic assembly,
and allows for intuitive design and understanding of systems.
The kinetic Tile Assembly Model (kTAM) is instead based on
the kinetics and thermodynamics of tiles in solution binding by
reversible bonds; this model serves as the basis for most of
what we discuss in this review.
In the aTAM, tile systems consist of non-rotatable square
tiles with up to four sticky ends of typically integer strengths:
usually 1 (‘‘single-strength’’ or ‘‘weak’’) or 2 (‘‘double-strength’’
or ‘‘strong’’). Growth begins from a designated seed tile and
subsequent tiles then attach asynchronously, one at a time.
Among all possible tiles, attachment is possible only if the tile’s
sticky ends match those on adjacent tiles in the assembly with
strengths that sum to at least the system ‘‘threshold’’ t. t is
often described as ‘‘temperature’’ in theoretical papers, but is
related physically to the relative favorability of attachment and
detachment, which could depend on a number of parameters,
including temperature, salt concentrations, and tile concentra-
tions. Once attached, tiles never detach.
Fig. 2 illustrates this process for a tile system at t = 2 that
implements exclusive-or (XOR) logic. As the system grows from
an initial seed tile at the bottom left, tiles with double-strength
Fig. 2 A t = 2 XOR/Sierpinski pattern tile system in the abstract Tile Assembly Model (aTAM). (a) Shows the 7 tiles in the set: a yellow seed tile with
double-strength ends (denoted by two lines), two boundary tiles, and four rule tiles. These are called the exclusive-or (XOR) rule tiles because the output
bonds are red if and only if the input bonds are different types. Tiles are colored according to their output bonds. (b) Shows a possible series of
attachments, with gray tiles indicating potential attachment sites. (c) Shows the possible tiles that can attach to a small assembly, while (d) shows a larger
17 17 assembly that would continue to grow.
Chem Soc Rev Review ArticlePublished on 10 May 2017. Downloaded by California Institute of Technology on 4/24/2026 7:33:00 PM.View Article Online
Page 4
This journal is © The Royal Society of Chemistry 2017 Chem. Soc. Rev., 2017, 46, 3808--3829 | 3811bonds form a V-shaped edge at the left and bottom, while tiles
implementing the XOR logic in the center attach by two single-
strength bonds: one from the tile left of the site, and one from
the tile below. The result is a Sierpinski triangle pattern that
grows out infinitely to the top right—a non-periodic fractal
pattern that is a consequence of the XOR function, and that
could not be produced by simple periodic tiling.
This cooperative attachment of tiles by two weak bonds
to two different tiles, where each bond is individually too
weak to allow attachment, makes growth at t = 2 particularly
interesting. t = 2 tile systems in the aTAM are capable of Turing-
universal computation.26,32 Squares33,34 and arbitrary shapes35
can be efficiently assembled by relatively few tile types, and
numerous other properties have been shown.11,12 These results
rely on cooperative attachment: with t = 1, at least for certain
classes of tile systems, significant computation cannot be
performed or is much more difficult.36
The aTAM is useful for considering the theoretical potential
of tile assembly, but ignores important aspects of physical
implementations. The kTAM derives from a different approach,
starting from basic physical considerations. DNA tiles, in general,
are monomers with ends that form reversible bonds to matching
sticky ends on other tiles. As tile systems involve large numbers of
tiles in solution, tile assembly can therefore be considered from
the perspective of chemical kinetics and thermodynamics for tile
attachment and detachment.
Here we focus on the growth of a single crystal in solution,
although later we will discuss a mass-action version of the kTAM
and other variants. To simplify the analysis for now, this most
basic version of the kTAM has a few fundamental assumptions:
(1) Free tiles are assumed to be in solution, at equal and
constant concentrations never depleted by assembly,
and to remain well-mixed, so diffusion can be ignored.
Thus growth of individual crystals will be independent of
the growth of other crystals in the solution.
(2) We only (for now) consider growth initiating from spe-
cific seed tiles, as in the aTAM, and only consider single
free tiles (monomers) attaching to and detaching from a
well-formed crystal.
(3) Tiles attach based only on diffusion. Since we are con-
sidering the growth of an individual crystal in solution,
attachment of a tile to an individual lattice site will
depend only on the concentration of the tile. In a marked
departure from the aTAM, we assume that bonds make
no difference here: the attachment rate will be the same
whether the tile attaches by one bond or three, or has a
bond that does not match. Tiles that could not attach in
the aTAM have the same chance of attaching as any
other tile.
(4) Once attached, tile detachment is assumed to depend
only on the total change in free energy for all the correct
bonds holding it to the lattice. We will assume that the
standard free energies of correct bonds can simply be
added (i.e. cooperative attachment does not increase or
decrease the favorability of the individual attachments)
and that mismatched bonds do not hinder binding.13
We will consider bonds in comparison to a reference standard
free energy of some typical single bond, DG
se, with se referring to
‘‘sticky end’’, as the term is commonly used in experimental
literature to refer to a bond. A tile will thus be attached by some
free energy bDG
se, where b is the sum of the ‘‘strengths’’ of all
correct bonds compared to DG
se. In many theoretical analyses, all
bonds are assumed to be identical, with strength 1 (i.e., DG
se),
or bonds are assumed to have some integer multiple strength
(e.g., DG
se or 2DG
se), but the kTAM does not require this, and
there have been some investigations of systems with differing
or sequence-dependent bond strengths.24
These assumptions result in the following rates for attach-
ment and detachment of a single tile at a single lattice site:
rf ¼ kf ½c rr;b ¼ kr;b ¼ kf u0ebDG
se=RTþa (1)
where kf is a forward rate determined experimentally, [c] is the
tile concentration, b is the total strength of the correct bonds
between the tile and adjoining tiles, u0 is a standard concen-
tration (1 M), and a is a constant unitless free energy change
from other factors, such as the loss of rotational entropy during
binding. These rates, illustrated in Fig. 3, describe the events
occurring on a single crystal grown from a seed.
In the kTAM literature, these rate equations are typically
used in an equivalent but more symmetric form. By defining
kˆf u0kfea, Gse DG
se
RT, and Gmc such that [c] = u0eGmc+a,
we can construct the symmetric rate equations
rf = kˆf eGmc rr,b = kˆf ebGse (2)
that are the core of the kTAM.26 In this form, kˆf is a reference
rate for tile dissociation in units of s1. Gse and Gmc are two
unitless free energy analogues. Gse is the sign-reversed, dimen-
sionless reference free energy of a typical single sticky end (se)
bond, where higher numbers correspond to stronger binding
and thus longer times before detachment. Gmc is a logarithmic
term for the free monomer concentration (mc), where larger
numbers correspond to exponentially lower concentrations and
longer times between attachments. Arranged in this manner,
rate equilibrium can be easily understood as a comparison
between Gmc and Gse: if, for example, Gmc is larger than Gse but
less than double Gse, then the attachment rate for a strength 2,
or b = 2 interaction, will be higher than its detachment rate, but
the attachment rate for a b = 1 interaction will not.
For a given tile set, seed tile, and values for Gmc and Gse, the
standard kTAM is a continuous-time Markov chain (CTMC) that
satisfies detailed balance. The initial state is an assembly
consisting only of the seed tile, and every subsequent step
involves the reversible addition or removal of a non-seed tile.
(As we are identifying the crystal by the seed, removal of the
seed tile itself is not considered.) Thus, considering a tile
addition to an assembly A that results in an assembly A0:
rf
rr;b
¼ ebGseGmc ¼ eDG (3)
where DG = G(A0) G(A) with respect to the energy of an
assembly, G(A) = NGmc BGse where N is the total number of
Review Article Chem Soc RevPublished on 10 May 2017. Downloaded by California Institute of Technology on 4/24/2026 7:33:00 PM.View Article Online
Page 5
3812 | Chem. Soc. Rev., 2017, 46, 3808--3829 This journal is © The Royal Society of Chemistry 2017tiles in the assembly and B is the summed strength of all
matching bonds in the assembly. Consequently, the equilibrium
probability of an assembly A, and the associated partition
function, conform to the Boltzmann distribution:
PðAÞ ¼ 1
Z eGðAÞ Z ¼ X
A
eGðAÞ: (4)
While this formula is convenient and can be insightful, it must
be used with caution because in many cases of interest the
CTMC state space is infinite, the partition function does not
converge, and there is no reachable equilibrium.
The kTAM employs assumptions that are not valid in
general: for example, tiles may have flexible structures that
reconfigure upon attachment, binding strength may not simply
be a sum of individual bond strengths, or experiments might
incorporate growth on a surface. However, results from DX tile
ribbon and nanotube growth in solution largely fit predictions
and simulations in the kTAM,37,38 and many experimental
results qualitatively fit the model.28,29,39 Departures from kTAM
assumptions often have effects that can be incorporated into
renormalized parameter values, and in experimental measure-
ments, these effects are often built in to measured values,
particularly Gse and a. For example, while directly imaging
growth on a surface with an atomic force microscope (AFM)
involves surface binding terms and an overall force from the
AFM tip, as approximately constant terms, these will both be
incorporated into a significantly different a in measurements,
as seen in experiments.13 More generally, so long as binding is
at least dependent primarily upon the total bond strength b and
some constant free energy of a bound tile, then Gse and a can be
chosen to make any two values of b at which the kTAM
detachment rate will be a good approximation, regardless of
nonlinear or constant effects. As discussed later, in many cases,
tile system growth in the kTAM is most interesting when tiles
attached by b o 1 fall off very quickly, and tiles attached by
b 4 2 fall off very slowly, with b = 1 and b = 2 involving the most
relevant rates. In this regime, even if there are significant
departures from the kTAM’s assumptions, so long as para-
meters cause the b = 1 and b = 2 cases to be approximately
correct, the kTAM will still give reasonable results. It is thus
hoped that such an approximation will be suitable for SSTs,
even though, as single strands with no rigid structure, they
must significantly change configuration during attachment.
Growth on a surface, not the focus of the kTAM, could also
involve multiple diffusion modes, with some tiles in solution
attaching directly to crystals, and others binding to the surface
itself and sliding in two dimensions. These two diffusion
modes could, for example, result in attachment rates depen-
dent upon lattice site location, with sites on the edges of
crystals having more frequent attachments than those in the
interior of crystals, and would require a different model. In
direct tile-level atomic force microscope observation of DX tile
crystal growth near equilibrium on a surface, this was not
found to be a significant factor. Single-tile attachments and
detachments rates fit the kTAM within error, but had a and Gse
values incorporating surface attachment and AFM effects that
were significantly different from typical values in solution.13
With other growth conditions or tile types, surface diffusion
could be significant: for example, surface diffusion effects have
been modelled and seen experimentally when using larger tile
structures made from T-shaped junctions or DNA origami.40–42
However, for the purposes of this review, we consider only
growth in solution.
Additionally, lattice defects and assembly–assembly interactions
have been observed, motivating designs and conditions to avoid
them,13,30 and nucleation and free tile depletion are notable
deviations discussed later in this review. Slightly different tile
structures, such as those with hairpin labels, have also been
found to potentially have significantly different energetics,
making their use in systems difficult.30,37
The kTAM can be applied to 1, 2, and 3D lattices. Most
theoretical analyses and simulations to date have not considered
3D lattices, though there has been some theoretical work on 3D
error-correcting schemes.43 Uniquely-addressed 3D structures
Fig. 3 The Sierpinski system from Fig. 2 in the kinetic Tile Assembly Model (kTAM). (a) Shows one possible path of attachments and detachments,
disregarding the time between each change. (b) Shows attachment and detachment rates for various tiles and lattice sites. (c) Shows the a larger
assembly, as in Fig. 2. Errors are shown in bold: at the top, a growth error is show, with errors indicating potential subsequent attachments that could
‘‘trap’’ the error in place in the kinetic trapping model, while at the bottom, a single error causes a different pattern to arise, despite later tiles (dotted)
attaching by correct bonds.
Chem Soc Rev Review ArticlePublished on 10 May 2017. Downloaded by California Institute of Technology on 4/24/2026 7:33:00 PM.View Article Online