of 12
ARTICLE
Received 7 Aug 2016
|
Accepted 21 Dec 2016
|
Published 23 Feb 2017
Compiler-aided systematic construction
of large-scale DNA strand displacement circuits
using unpurified components
Anupama J. Thubagere
1
, Chris Thachuk
2
, Joseph Berleant
2
, Robert F. Johnson
1
, Diana A. Ardelean
3
,
Kevin M. Cherry
1
& Lulu Qian
1,2
Biochemical circuits made of rationally designed DNA molecules are proofs of concept for
embedding control within complex molecular environments. They hold promise for trans-
forming the current technologies in chemistry, biology, medicine and material science by
introducing programmable and responsive behaviour to diverse molecular systems. As the
transformative power of a technology depends on its accessibility, two main challenges are an
automated design process and simple experimental procedures. Here we demonstrate the
use of circuit design software, combined with the use of unpurified strands and simplified
experimental procedures, for creating a complex DNA strand displacement circuit that
consists of 78 distinct species. We develop a systematic procedure for overcoming the
challenges involved in using unpurified DNA strands. We also develop a model that takes
synthesis errors into consideration and semi-quantitatively reproduces the experimental data.
Our methods now enable even novice researchers to successfully design and construct
complex DNA strand displacement circuits.
DOI: 10.1038/ncomms14373
OPEN
1
Bioengineering, California Institute of Technology, 1200 East California Boulevard, Pasadena, California 91125, USA.
2
Computer Science, California Institute of
Technology, 1200 East California Boulevard, Pasadena, California 91125, USA.
3
Applied and Computational Mathematics, California Institute of Technology, 1200
East California Boulevard, Pasadena, California 91125, USA. Correspondence and requests for materials should be addressed to L.Q. (email: luluqia
n@caltech.edu).
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
1
T
he success of computer engineering has inspired attempts
to use hierarchical and systematic approaches for devel-
oping molecular devices with increasing complexity. To
enable the design and construction of a wide range of functional
molecular systems, we need software tools such as a compiler that
can automatically translate high-level functions to low-level
molecular implementations and provide models and simulations
for predicting and debugging the behaviours of designed
molecular systems. The mechanism of DNA strand displacement
has been used to create a variety of synthetic molecular systems
including circuits, motors and triggered assembly of structures
1
.
Software tools have been developed for designing and analysing
DNA strand displacement systems, capable of generating nucleic
acid sequences from well-defined structures and molecular
interactions
2,3
, calculating the thermodynamic
2,4,5
and kinetic
6
properties of designed molecules, and evaluating if the behaviours
of the molecular systems agree with the higher-level designs
3,7–11
.
There also exist a few molecular compilers that can translate
abstract functions such as a logic function to DNA strand
displacement implementations without requiring an under-
standing of the molecular level details
12,13
. However, there
has been little independent experimental validation of these
compilers, most of which were developed in parallel with or after
experimental findings
12,14
.
In addition to software tools that facilitate automated design
and analysis of DNA strand displacement circuits, we also need to
simplify the experimental procedures for creating these circuits
in vitro
, so that it is possible for researchers with diverse
backgrounds to build their own circuits and explore potential
applications. A great inspiration is DNA origami
15
, a technique
that folds DNA into sophisticated structures. In just 10 years
since its birth, DNA origami has become one of the most
significant successes in the field of DNA nanotechnology.
Over 170 research groups have contributed to advancing this
technique or developing it for applications in a variety of
research areas
16–19
. A fundamental reason why DNA origami
was able to quickly spread around the world is that the
experimental procedure is extremely simple and makes use of
cheap, unpurified nucleic-acid strands. In contrast, other
than a few very simple circuits with just one or two double-
stranded components
20
, most DNA strand displacement circuits
were constructed using strands that were purchased either
purified or unpurified, but all followed by in-house
polyacrylamide gel electrophoresis (PAGE) purification to
reduce undesired products due to synthesis errors and
stoichiometry errors
12,14,21
. Purified strands are approximately
ten times more expensive than unpurified strands, which
significantly increases the cost for building large-scale DNA
circuits. In-house PAGE purification is both time consuming and
labour intensive.
In this work, we show that one can successfully build
a complex DNA strand displacement circuit, using DNA
sequences automatically generated from a molecular compiler.
We also show that one can even do so using cheap, unpurified
DNA strands, following simple and systematic experimental
procedures.
Results
Circuit design
. A simple DNA strand displacement motif called
the seesaw gate was developed to scale up the complexity of
DNA circuits
22
and was used to demonstrate digital logic
computation
12
and neural network computation
23
. The Seesaw
Compiler
12,24
was developed to automatically translate an
arbitrary feed forward digital logic circuit into its equivalent
seesaw DNA circuit (Fig. 1). The compiler takes an input file that
describes a logic circuit with a list of input and output terminals,
and a list of AND, OR, NOT, NAND and NOR gates with the
connectivity of their terminals specified. First, a technique called
dual-rail logic is applied to translate the original logic circuit into
an equivalent circuit that contains AND and OR gates only
25
.
This is because the NOT gate cannot be directly implemented in
multi-layer use-once DNA circuits, if the OFF and ON state of a
signal is represented by low and high concentration of a single
DNA strand, respectively. If a NOT gate were implemented this
way, then output molecules of the gate could be immediately
produced in the absence of input. However, once this reaction
reaches equilibrium it cannot be reversed, even if input mole-
cules are added at a later point. With dual-rail logic, each terminal
in the original circuit is replaced by two terminals, representing
the OFF and ON states of a signal separately (for example,
each input signal
x
i
is replaced by
x
0
i
and
x
1
i
). Thus, no reaction
will take place until signal molecules on one of the two wires
have arrived. With this representation, the NOT gate
can be implemented by exchanging the two wires of an
input and output signal. Each AND, OR, NAND and NOR
gate in the original circuit is replaced by a pair of AND and OR
gates.
Next, the compiler translates the dual-rail logic circuit into an
equivalent seesaw DNA circuit. In a seesaw DNA circuit, each
signal is defined as a wire
w
j
,
i
connecting seesaw nodes
j
and
i
,
and implemented using a single-stranded DNA molecule.
Each AND and OR gate in the dual-rail circuit is replaced by a
seesaw AND and OR gate, respectively, which is defined as a
pair of integrating and amplifying seesaw nodes connected with a
set of input and output wires
12
. The seesaw nodes are
composed of double-stranded threshold and gate:output
molecules and single-stranded fuel molecules (Fig. 1, bottom
right). We will explain how the seesaw logic gates work in the
next section. Input fan-out gates are introduced to take an input
signal that is used for multiple logic gates and produce the
corresponding number of output signals. Reporters are
introduced to take each output signal and generate a distinct
fluorescence signal for readout.
Finally, the compiler generates Visual DSD
3,26
code and
Mathematica code for simulating and analysing the seesaw DNA
circuit and a file that contains DNA sequences for all molecular
species in the circuit. The Visual DSD code can be used to
automatically produce diagrams of species, reactions and network
graphs with domain-level representation of DNA and to simulate
the circuit behaviour based on the network of chemical reactions.
The Mathematica code provides more customized and efficient
simulations of seesaw circuits. The simulation uses the
CRNSimulator package
27
and models a specific set of side
reactions in addition to the designed reactions in a seesaw
network
12
.
As a demonstration of using the Seesaw Compiler, we
designed a single DNA strand displacement circuit that
implements two distinct elementary cellular automata
transition functions. An elementary cellular automaton (CA) is
one of the simplest models of computation
28
. It consists of a
one-dimensional grid of cells, collectively called a generation,
where each cell has a binary state of 0 or 1. In each subsequent
generation, the state for a cell
C
is determined by its current
state and those of its left neighbour
L
and right neighbour
R
. A state transition rule maps each of the 2
3
¼
8 possible
combinations of states for
L
,
C
and
R
to either 0 or 1.
Thus, a length 8 binary string uniquely identifies one of the 2
8
possible transition functions that specify how an elementary
CA will evolve between generations. The rule 110 elementary CA
(binary number 01101110 written in decimal) is famously known
to be Turing universal
29
.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14373
2
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
Another rule that is equally powerful is rule 124 (binary
number 01111100 written in decimal), generated by applying the
following mirror transformation: the new state of the centre cell
for
LCR
¼
zyx
in rule 124 is the same as the new state for
LCR
¼
xyz
in rule 110. Our circuit was designed to compute a
combined logic function of the two transition rules (Fig. 2a). It
consists of five logic gates in two layers, including a three-input
two-output NAND gate. It is noteworthy that we designed the
circuit to demonstrate an interesting logic function associated
with cellular automata and not to implement the actual cellular
automata model. The circuit operates in a well-mixed test tube
environment that does not involve spatial dynamics (that is, no
geometry of cells).
The DNA circuit generated by the Seesaw Compiler consisted
of 6 layers and a total of 78 distinct initial DNA species (Fig. 2b
and Supplementary Fig. 1). Mathematica simulations of the DNA
INPUT(1)
# x1
INPUT(2)
# x2
INPUT(3)
# x3
INPUT(4)
# x4
OUTPUT(11)
# y1
OUTPUT(8)
# y2
5 = NOR(1, 2)
6 = NOT(4)
7 = AND(3, 6)
8 = OR(3, 4)
9 = NAND(5, 3, 4)
10 = OR(5, 7)
11 = NAND(9, 10)
Feedforward logic circuit
INPUT(2)
# x1^0
INPUT(3)
# x1^1
OUTPUT(22)
# y1^0
OUTPUT(23)
# y1^1
10 = OR(3, 5)
11 = AND(2, 4)
14 = OR(6, 9)
15 = AND(7, 8)
16 = AND(6, 8)
17 = OR(7, 9)
18 = AND(11, 7, 9)
19 = OR(10, 6, 8)
INPUT(2) = w[5,22]
# x1^0
INPUT(3) = w[7,20]
# x1^1
OUTPUT(22) = Fluor[52] # y1^0
OUTPUT(23) = Fluor[54] # y1^1
inputfanout[13,12,{28,32,38}]
inputfanout[15,14,{30,34,36}]
seesawOR[20,21,{7,11},{38,40}]
seesawAND[22,23,{5,9},{36,42}]
seesawOR[28,29,{13,19},{40}]
seesawAND[30,31,{15,17},{42}]
Reporter[52,45]
Reporter[54,47]
Seesaw DNA circuit
directive plot <_ _ _ Fluor52> (* y1^0 *)
directive plot <_ _ _ Fluor54> (* y1^1 *)
def normal = 0.0003 (* normal toehold binding rate constant nM^-1 s^-1*)
def slow = 0.000015 (* slow toehold binding rate constant nM^-1 s^-1*)
(* a seesaw signal *)
def signal(N,iL,i,iR,jL,j,jR) = ( N * <iL^ iiR^ T^ jL^ j jR^>
)
(* 2-input 2-output seesaw OR gate *)
def seesawOR2I2O(i1L,i1,i1R,i2L,i2,i2R,k1L,k1,k1R,k2L,k2,k2R)
=
(gateL(20*N,i1L,i1,i1R,i2L,i2,i2R)
| thresholdL(6*N,i1R,i2L,i2,i2R)
| gateL(10*N,i2L,i2,i2R,k1L,k1,k1R)
| gateL(10*N,i2L,i2,i2R,k2L,k2,k2R)
| signal(40*N,i2L,i2,i2R,fL,f,fR))
( signal(ON,S5L,S5,S5R,S22L,S22,S22R) (* x1^0 *)
| signal(OFF,S7L,S7,S7R,S20L,S20,S20R) (* x1^1 *)
| seesawOR2I2O(S20L,S20,S20R,S21L,S21,S21R,S38L,S38,S38R,S40L,S40,S40R)
| seesawAND2I2O(S22L,S22,S22R,S23L,S23,S23R,S36L,S36,S36R,S42L,S42,S42R)
Visual DSD code
(* Rate constants: *)
kf = 2*10^6; (* fast strand displacement rate, unit: M^ –1 s^ –1 *)
ks = 5*10^4; (* slow strand displacement rate, unit: M^ –1 s^ –1 *)
(* Translates a seesaw gate into a list of reactions: *)
seesaw[x_,l_List,r_List]:={
(* Toehold exchange reactions *)
Outer[revrxn[w[#1,x]+g[x,w[x,#2]],g[w[#1,x],x]+w[x,#2],ks,ks]&,l,r],
(* Translates logic OR operation into a list of seesaw gates *)
seesawOR[x1_,x2_,l_List,r_List]:=Module[{f},
{seesaw[x1,l,{x2}],
(* Simulation *)
SIMcircuit=Table[gatesys={
seesawOR[20,21,{7,11},{38,40}],
seesawAND[22,23,{5,9},{36,42}],
(* Plot *)
Plot[Evaluate[SIMcircuit],{t,0,time},
Mathematica code
x1^0: w5,22 = S22 T S5
x1^1: w7,20 = S20 T S7
Th12,13:13-t = S13
Th12,13:13-b = s12* T* S13*
w13,28
= S28 T S13
w13,32
= S32 T S13
G13-b
= T* S13* T*
w13,f
= Sf T S13
Rep48-t
= RQ S48
Rep48-b
= T* S48* ATTO590
DNA sequences
Visual DSD simulation
Mathematica simulation
+
+
+
+
53
f
5
2
1
6
Gate:Output
(
G
5:5,6
)
Fuel (
w
5,f
)
Reporter (
Rep
6
)
Input (
w
53,5
)
Threshold (
Th
53,5:5
)
S41L
S41
S41R
S41L*
S41*
S41R*
T*
S40R*
S41L
S41
S41R
T
S40R
S40
S40L
S41L
S41
S41R
S41L
S41
S41R
T
S40R
S40
S40L
S41L*
S41*
S41R*
T*
S40R*
S41L
S41
S41R
S42L
S42 S42R
T
S41L*
S41*
S41R*
T*
T*
S41L
S41
S41R
T
S40R
S40
S40L
S41L
S41
S41R
S42L
S42
S42R
T
S41L*
S41*
S41R*
T*
T*
S41L
S41
S41R
T
S40R
S40
S40L
Dual-rail logic circuit
1.0
y
2
0
y
2
1
y
1
0
y
1
1
x
4
x
3
x
2
x
1
= 1001
0.8
0.6
0.4
0.2
0.0
0
2
4
6
8
10
Time (hours)
Output
S5
S5
S5
Sf
S5
T
S5
*
s53
*
T
*
S6
Q
F
S53
S5
*
T
*
S6
*
T
*
T
*
S6
T
T
–0.5
–1.5
1
S5 =
S7 =
S9 =
S5* =
S7* =
S9* =
Figure 1 | Automated circuit design steps using the Seesaw Compiler.
A feedforward digital logic circuit is first translated into an equivalent dual-rail logic
circuit and then translated into an equivalent seesaw DNA circuit. Visual DSD code and Mathematica code are generated for analysing and simulating th
e
seesaw DNA circuit, and DNA sequences are generated for constructing the circuit. Bottom right diagram introduces the notations of seesaw circuits:
black
numbers indicate identities of nodes. The locations and values of red numbers indicate the identities of distinct DNA species and their relative init
ial
concentrations, respectively.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14373
ARTICLE
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
3
circuit predicted correct computation for all 8 possible input
combinations under ideal experimental conditions (Fig. 2c).
The next step was to construct the DNA circuit using
strands that were purchased unpurified and with no additional
in-house purification. We expected that the main challenges
would be to understand how synthesis errors and stoichiometry
errors affect the behaviours of DNA circuits and to
explore solutions that restore the desired circuit behaviour.
We took a bottom-up approach and began building the DNA
circuit from the simplest functi
onal component—digital signal
restoration.
Calibrating effective concentrations
. Digital signal restoration is
a process that pushes the intrinsically analog signal towards either
the ideal ON or OFF state, therefore cleaning up the noise and
compensating for the signal decay that occurs during circuit
execution. In seesaw circuits, digital signal restoration is a
component of every logic gate, and is implemented by an
amplifying seesaw node with the following idealized input-output
function:
y
¼
1
x
4
th
0
x

th

ð
1
Þ
At the molecular level, the digital signal restoration process
consists of two basic reactions: catalysis and thresholding.
Catalysis is implemented with two toehold exchange pathways
that release free output strands
w
i
,
k
from double-stranded gate
molecules
G
i
:
i
,
k
, using the input strands
w
j
,
i
as a catalyst (Supple-
mentary Fig. 2a):
w
j
;
i
þ
G
k
s
i
:
i
;
k
!
w
j
;
i
þ
w
i
;
k
ð
2
Þ
LCR
= 100
LCR
= 001
LCR
= 101
LCR
= 110
LCR
= 111
0 0 0
0
0
0 0 1
0
1
0 1 0
1
1
0 1 1
1
1
1 0 0
1
0
1 0 1
1
1
1 1 0
1
1
1 1 1
LCR
= 000
0
0
Rule 110
Rule 124
LCR
= 010
LCR
= 011
L
C
R
L C R
L
0
L
1
C
0
C
1
R
0
R
1
R
124
0
R
124
1
R
110
1
R
110
0
R
124
R
124
R
110
R
110
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
1.0
0.8
0.6
0.4
0.2
0.0
0246810
Time (h)
Output
R
124
0
R
124
1
R
110
0
R
110
1
ab
c
Figure 2 | Design of a rule 110–124 circuit using the Seesaw Compiler.
(
a
) Gate diagram and truth table of a digital logic circuit that computes the
transition rules 110 and 124 of elementary cellular automata. (
b
) Seesaw gate diagram of the equivalent DNA strand displacement circuit. Each seesaw
node connected to a dual-rail input implements input fan-out. Each pair of seesaw nodes labelled
4
and
3
implements a dual-rail AND and OR gate,
respectively. Each pair of dual-rail AND and OR gates implements an AND, OR or NAND gate in the original logic circuit. Each dual-rail output is convert
ed
to a fluorescence signal through a reporter, indicated as a half node with a zigzag arrow. Each circle and dot inside a seesaw node indicates a double-
stranded threshold and gate molecule, respectively. Each dot on a wire indicates a single-stranded fuel molecule. (
c
) Simulations of the DNA strand
displacement circuit using the previously developed model for purified seesaw circuits. Trajectories and their corresponding outputs have matchin
g colours.
Overlapping trajectories were shifted to be visible. Dotted and solid lines indicate dual-rail outputs that represent logic OFF and ON, respectivel
y. For
example, when input
LCR
¼
001, meaning
L
0
,
C
0
and
R
1
were introduced at a high concentration and
L
1
,
C
1
and
R
0
at a low concentration, two output
trajectories
R
124
0
and
R
110
1
reached an ON state and the other two output trajectories
R
124
1
and
R
110
0
remained in an OFF state, indicating that the output
was computed to be 0 and 1 for rule 124 and 110, respectively. Simulations were performed at 1
50 nM—the compiler recommended standard
concentration for large-scale purified seesaw circuits.
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14373
4
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
Catalysis can be used for signal amplification, since a small
amount of input can trigger the release of a much larger amount
of output.
Thresholding is implemented with double-stranded threshold
molecules
Th
j
,
i
:
i
consuming the input at a much faster rate
(
k
f
c
k
s
) than the input acting as a catalyst (Supplementary
Fig. 2b):
w
j
;
i
þ
Th
j
;
i
:
i
k
f
!
+
ð
3
Þ
Asshowninsimulationsge
nerated using the Seesaw
Compiler (Fig. 3a), when the concentration of the
threshold molecule is 0.5

(where 1

is a standard
concentration of 100 nM), we expect that input less
than the threshold (for example, 0.3

)shouldbecleaned
up to an ideal OFF state via reaction 3 and input greater
than the threshold (for example, 0.7

) should be amplified
to an ideal ON state via reaction 2. However, the
observed circuit behaviour
was different: when input
¼
0.7

,
the output signal was higher than an ideal OFF state, but
did not reach an ideal ON state (
Fig. 3b). This experimental
result suggested that the input did not sufficiently exceed
the threshold, which was an indication that the effective
concentration of an unpurified
threshold species, compared
with that of an unpurified signal species, was higher than
expected.
The nominal concentration of a DNA species can be
measured using ultraviolet absorbance, but it can be higher
than the effective concentration, which is the concentration of
the DNA species actually perform
ing the desired reactions. If
the sequences of the DNA strands are properly designed, the
difference between nominal concentration and effective con-
centration is typically caused by synthesis errors including
nucleotide insertion, deletion and mismatch. To calibrate the
effective concentrations of unpurified DNA molecules, we
defined the following ratio between effective (eff) and nominal
(nom) concentrations of an arbit
rary signal, threshold and gate
species:
a
j
;
i
¼
w
j
;
i

eff
w
j
;
i

nom
ð
4
Þ
b
j
;
i
¼
Th
j
;
i
:
i

eff
Th
j
;
i
:
i

nom
ð
5
Þ
g
i
;
k
¼
G
i
:
i
;
k

eff
G
i
:
i
;
k

nom
ð
6
Þ
The effective to nominal concentration of a DNA species
cannot be measured in isolation. More importantly, the absolute
values of
a
,
b
and
g
should only affect the speed but not the
correctness of computation, if the values remain comparable to
each other. Thus, we chose to estimate the ratio between
b
and
a
for a threshold consuming a signal, by comparing simulation
with experimental result of a signal restoration circuit. For
example, manipulating the threshold value in simulation (sim)
identified that
Th
53
;
5
:
5

sim
¼
0
:
7

agreed with the experimental
data (Fig. 3c), which means the effective concentration of the
threshold was similar to that of the signal for
Th
53
;
5
:
5

nom
¼
0
:
5

and
w
53
;
5

nom
¼
0
:
7

. Thus, the threshold to signal ratio can be
calculated as:
b
53
;
5
a
53
;
5
¼
Th
53
;
5
:
5

eff
Th
53
;
5
:
5

nom

w
53
;
5

nom
w
53
;
5

eff
¼
w
53
;
5

nom
Th
53
;
5
:
5

nom





Th
53
;
5
:
5
½
eff
¼
w
53
;
5
½
eff
¼
0
:
7
0
:
5
¼
1
:
4
ð
7
Þ
A possible explanation for an unpurified threshold having a
higher effective concentration than an unpurified signal, when the
nominal concentrations are the same, is the following: the
synthesis errors of an unpurified strand depend on the length of
the strand, because in the process of chemical synthesis each
nucleotide is attached to a growing chain of oligonucleotide
one at a time and the coupling efficiency of each step is less than
a
53
f
5
2
2
x
1
x
1
x
2
x
1
x
2
x
1
x
2
x
1
x
1
x
2
x
2
x
2
18
22
1
6
10
f
1
2
2
x
1
x
2
y
y
y
yy
21
27
1
23
53
f
5
2
Input
W
53,5
1
6
OR
10
1
1
y
23
–1.5
AND
Input
0
0.3
0.7
1
Input
0
0.3
0.7
1
Input
0
0.3
0.7
1
0
0
1
1
0
1
0
1
00
0
1
2
0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Time (h)
0246810
Time (h)
0246810
Time (h)
0246810
Time (h)
0246810
Time (h)
0246810
Time (h)
02468
[
Th
53,5:5
]
nom
= 0.85 ×
[
Th
10,1:1
]
nom
= 0.35 ×
–0.6
–1.5
–1.2
–1.5
[
G
1:1,23
]
tri
= 0.8 ×
[
Th
53,5:5
]
sim
= 0.5 ×
[
Th
53,5:5
]
nom
= 0.5 ×
[
Th
53,5:5
]
sim
= 0.7 ×
[
Th
53,5:5
]
nom
= 0.5 ×
–0.5
–1.5
bc
d
e
Figure 3 | Calibrating effective concentrations.
(
a
) Simulations and (
b
) experimental data of digital signal restoration. (
c
) Estimating effective threshold
concentration by fitting simulations to the data obtained. (
d
) OR and AND logic gates constructed using adjusted nominal threshold concentrations.
(
e
) Estimating effective gate concentration. Data show steady-state fluorescence level. 1
100 nM. Here and in later figures, all output signals in the
data were normalized using the minimum fluorescence signal (the first data point) of an OFF trajectory as 0 and the maximum fluorescence signal
(the average of the last five data points) of an ON trajectory as 1.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14373
ARTICLE
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
5
100% (ref. 30). Threshold mol
ecules are composed of shorter
strands (15 and 25 nucleotides) than signal molecules
(33 nucleotides) and thus may contain fewer synthesis errors.
Additional signal restoration experiments suggested that the
threshold to signal ratio
b
/
a
¼
1.4 was consistent for different
threshold and signal molecules (Supplementary Fig. 3). Thus,
using this ratio, we can then calculate how to adjust the nominal
thresholds for correctly computing logic AND and OR.
Each seesaw logic gate has an integrating node upstream of an
amplifying node. Ideally, an integrating node outputs the sum of
all inputs:
y
¼
X
n
i
¼
1
x
i
ð
8
Þ
A two-input logic function can be computed as:
y
¼
1
x
1
þ
x
2
4
th
0
x
1
þ
x
2

th

ð
9
Þ
Assuming that an ideal OFF state is [0, 0.2] and an ideal ON state
is [0.8, 1],
th
¼
0.6 will compute logic OR and
th
¼
1.2 will
compute logic AND, if the effective concentrations of the
threshold and input signals are comparable to each other (that
is,
b
/
a
¼
1).
As
b
/
a
a
1 for unpurified threshold and signal molecules, we
can take this ratio into consideration while calculating the lower
and upper bounds of the nominal threshold for an
n
-input logic
gate:
0
:
2
n

a
b

Th
OR
½
nom
o
0
:
8

a
b
ð
10
Þ
n

1
ðÞþ
0
:
2
½
a
b

Th
AND
½
nom
o
0
:
8
n

a
b
ð
11
Þ
Using
b
/
a
¼
1.4, we chose a nominal threshold of 0.35

and
0.85

for two-input OR and AND gate, respectively, and
0.4

and 1.6

for three-input OR and AND gate. Experiments
of the logic gates showed desired behaviours (Fig. 3d and
Supplementary Fig. 4).
An alternative approach for adjusting the nominal threshold is
to use the following equations:
Th
OR
½
nom
¼
0
:
6

a
b
ð
12
Þ
Th
AND
½
nom
¼
n

1
ðÞþ
0
:
2
½
a
b
ð
13
Þ
Compared with choosing a nominal threshold based on
the lower and upper bounds, this approach is less flexible but
simpler.
Next, we can estimate the ratio between
g
and
a
for a gate
releasing a signal, using an ex
periment that compares the
fully triggered (tri) concentration of the gate with the signal
when their nominal concentrat
ions are the same. For example,
thedatainFig.3eshowedthat
G
1
:
1
;
23

tri
¼
0
:
8

when
G
1
:
1
;
23

nom
¼
w
1
;
23

nom
¼
1

.Thus,thegatetosignalratiocan
be calculated as:
g
1
;
23
a
1
;
23
¼
G
1
:
1
;
23

eff
G
1
:
1
;
23

nom

w
1
;
23

nom
w
1
;
23

eff
¼
G
1
:
1
;
23

eff
w
1
;
23

eff





G
1
:
1
;
23
½
nom
¼
w
1
;
23
½
nom
¼
0
:
8
1
¼
0
:
8
ð
14
Þ
Additional gate calibration experiments suggested that the ratio
g
/
a
¼
0.8 was consistent for different gate and signal molecules
(Supplementary Fig. 5). We suspect that due to synthesis errors in
gate molecules, not all gates can successfully release a signal,
which is why an unpurified gate has a lower effective concen-
tration compared to a signal.
As signal restoration was built in within every logic gate to
accept an ON state of [0.8, 1], we decided not to make any
adjustment for nominal gate concentrations if
g
/
a
Z
0.8. Other-
wise, nominal concentration of an amplifying gate and an
n
-input
integrating gate can be adjusted as:
G
AMP
½
nom
¼
1

a
g
ð
15
Þ
G
INT
½
nom
¼
n

a
g
ð
16
Þ
Importantly, the values of
a
,
b
and
g
should depend on the
strand quality and thus could vary with different DNA synthesis
providers, procedures and even batches. It is necessary to
recalculate the ratios
b
/
a
and
g
/
a
, if these conditions change.
Identifying outliers
. With calibrated logic gates, we investigated
how well they compose together in larger circuits. We constructed
a two-layer logic circuit that is part of the rule 124 sub-circuit and
is composed of an AND gate and two upstream OR gates
(Fig. 4a). The expected circuit behaviour is that the output should
remain OFF when only one of the upstream OR gates is ON.
However, the observed circuit behaviour showed that the output
was reasonably OFF when one upstream OR gate was ON, but
was half ON when the other upstream OR gate was ON. This
experimental result suggested that the ON signals pushed onto
the two input wires of the downstream AND gate (that is, the
output wires of the two upstream OR gates) were significantly
different from each other, which was an indication that the
effective concentrations of the two unpurified gate species that
released the output signals were different—one of the gates must
be an outlier with
g
/
a
a
0.8.
Indeed, with a gate calibration experiment shown in Fig. 4b,
we measured that
g
18,53
/
a
18,53
¼
0.8

for one gate and
g
22,53
/
a
22,53
¼
0.44

for another. A possible explanation is that
the synthesis errors of unpurified strands somewhat depend on
DNA sequences
30
and variations of effective concentrations may
occur between different gate or threshold species. We suspect it
was not a coincidence that the outlier gate had a lower effective
concentration compared with other unpurified gates, because a
particular DNA strand having much worse quality than average is
probably more likely than it having much better quality.
Once an outlier is identified, either a threshold or a gate, the
nominal concentration can be adjusted using its own threshold to
signal ratio (that is,
b
/
a
) or gate to signal ratio (that is,
g
/
a
), the
common nominal concentration described in the previous
section, and the common ratio for other thresholds and gates:
Th
j
;
i
:
i

0
nom
¼
Th
j
;
i
:
i

nom

a
j
;
i
b
j
;
i

b
a
ð
17
Þ
G
i
:
i
;
k

0
nom
¼
G
i
:
i
;
k

nom

a
i
;
k
g
i
;
k

g
a
ð
18
Þ
We constructed the two-layer logic circuit using the adjusted
nominal gate
G
22
:
22
;
53

0
nom
¼
1
=
0
:
44

0
:
8
¼
1
:
8

(Fig. 4c). The
trajectories that compute logic ON reached an ideal high
fluorescence state faster than the previous experiments shown
in Fig. 4a and the trajectories that compute logic OFF remained at
a lower fluorescence state that were roughly identical for all three
input combinations, regardless of which upstream OR gate was
ON. However, after identifying and adjusting the outlier gate, we
still had a problem: the OFF trajectories were not at an ideal low
ARTICLE
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14373
6
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
fluorescence state. This led to the next tuning step that is
necessary for unpurified seesaw circuits.
Tuning circuit output
. Comparing the behaviour of the AND
gate when it was in isolation (Fig. 3d) and that when it was
connected with two upstream OR gates (Fig. 4c), the ON/OFF
separation was significantly decreased in the latter. These
experimental results suggest that, compared with purified seesaw
DNA circuits in which the ON/OFF separations were roughly
identical from a single logic gate to four-layer logic circuits
12
,
unpurified circuits are much noisier and the behaviour becomes
less robust with more than one layer. We suspect this is caused by
the stoichiometry errors in unpurified gate species. The double-
stranded gate molecules were annealed with the same amount of
top and bottom strands, because both strands have combinations
of toehold and branch migration domains that can cause
undesired interactions with other circuit components and thus
neither should be in excess. However, due to variations in the
pipetting volume and in the accuracy of concentrations, the equal
stoichiometry cannot be guaranteed. Without purification, a
small excess of one strand or another in the gate species cannot be
removed. Therefore, the excess of strands would result in
undesired release of output signals in logic gates, even without
input signals, and introduce extra noise to downstream logic
gates.
Fortunately, thanks to the thresholding function in every logic
gate, we can tune the circuit output by increasing a threshold.
A simple method for estimating how much threshold adjustment
is needed is based on the ON/OFF separation of the circuit
output. Using experimental data of a logic circuit with different
inputs, we can choose a trajectory that should compute logic ON
and OFF, respectively, and calculate the difference (
d
) between
the observed OFF value and an ideal OFF value, when the ON
trajectory reaches an ideal ON value. Considering 0.7 and 0.3 as
the lower bound, and 0.9 and 0.1 as the upper bound for an ideal
ON/OFF separation, the range of
d
can be determined as:
y
OFF


y
ON
¼
0
:
7

0
:
3

d

y
OFF


y
ON
¼
0
:
9

0
:
1
ð
19
Þ
The nominal threshold in the logic gate that produces the circuit
output can then be adjusted accordingly:
Th
j
;
i
:
i

0
nom
¼
Th
j
;
i
:
i

nom
þ
d

a
b
ð
20
Þ
Using the data of the two-layer logic circuit shown in Fig. 4c,
we chose the trajectory with input
¼
01010 and 11100 as the
reference ON and OFF trajectory, respectively, and calculated
0.08
r
d
r
0.41. We then increased the threshold in the down-
stream AND gate to
Th
53
;
5
:
5

0
nom
¼
0
:
85
þ
0
:
28
=
1
:
4
¼
1
:
05

and
repeated the experiment. The circuit behaviour was improved
with a much better ON/OFF separation (Fig. 5a).
With the same method, we constructed another two-layer logic
circuit that is composed of an OR gate and two upstream AND
gates (Fig. 5b). In this case, using input
¼
00011 and 01110 as the
reference ON and OFF trajectories, we obtained a similar range of
d
and decided to apply the same amount of increase to the
threshold in the downstream OR gate.
It is noteworthy that a rule of thumb is to choose the slowest
ON trajectory and the fastest OFF trajectory as the references for
threshold adjustment, but different choices can be made if one
has the knowledge of which data set is experimentally more
reliable. Also note that increasing the threshold not only
suppresses the OFF trajectories but also slows down the ON
trajectories and thus this method of tuning the circuit output is
only applicable if all ON trajectories are significantly faster than
all OFF trajectories (which should be true if the thresholds and
gates are properly calibrated).
Combining the two logic circuits shown in Fig. 5 and adding
fan-out gates for input signals that are used in multiple logic
gates, we successfully demonstrated the rule 124 sub-circuit
consisting of 54 distinct DNA species (Supplementary Fig. 6).
a
bc
ON
OFF
ON
OFF
18
f
53
2
12
6
y
y
–1.5
–1.5
34
x
1
x
3
x
2
x
3
22
f
53
2
12
6
39
53
f
y
5
2
21
6
39
f
22
2
2
29
35
1
34
f
18
2
3
28
37
–0.6
1
33
1
1
0
1
1
0
1
0
0
0
0
1
0
0
1
1
1
0
0
0
1
0
y
0
2
0
2
1
1
0
1
1
0
0
1
0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Time (h)
0246810
Output
1.0
0.8
0.6
0.4
0.2
0.0
Output
1.0
0.8
0.6
0.4
0.2
0.0
Time (h)
x
1
1
1
0
1
0
1
0
0
x
2
1
1
1
0
1
0
0
0
x
3
0
1
0
0
1
1
0
1
x
4
0
0
1
0
1
0
1
1
x
5
0
1
1
0
1
1
1
0
y
1
0
1
1
0
0
1
0
x
1
1
1
0
1
0
1
0
0
x
2
1
1
1
0
1
0
0
0
x
3
0
1
0
0
1
1
0
1
x
4
0
0
1
0
1
0
1
1
x
5
0
1
1
0
1
1
1
0
y
Time (h)
0123
0246810
S53
G
22:22,53
G
22:22,53
G
18:18,53
G
18:18,53
[
G
22:22,53
]
tri
= 0.44 ×
[
G
22:22,53
]
nom
= 1.8 ×
[
G
18:18,53
]
tri
= 0.8 ×
S53
S18
S18
T
*
T
*
*
S22
T
*
T
*
*
T
T
S22
S53
S22
T
*
T
*
*
T
S22
S53
S18
T
*
T
*
*
T
S18
x
1
x
2
x
3
x
1
x
2
x
3
x
4
x
5
x
1
x
2
x
3
x
4
x
5
[
Th
53,5:5
]
nom
= 0.85 ×
[
Th
34,18:18
]
nom
= 0.4 ×
[
Th
39,22:22
]
nom
= 0.35 ×
–1.2
–1.5
–0.6
Figure 4 | Identifying an outlier gate.
(
a
) Logic circuit diagram, seesaw circuit diagram and experimental data of a two-layer logic circuit. (
b
) Measuring
the effective concentrations of the gate species. Three independent circuits were used to measure the effective concentrations of two gates fully tr
iggered
by
x
1
and
x
2
, respectively, comparing with the effective concentration of
x
3
(using signal strand
w
18,53
). (
c
) Experimental data of the two-layer logic circuit
using adjusted nominal gate concentration. 1
100 nM.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14373
ARTICLE
NATURE COMMUNICATIONS
| 8:14373 | DOI: 10.1038/ncomms14373 | www.nature.com/naturecommunications
7