LETTERS

NATURE BIOTECHNOLOGY

original library before inserting the DNA into the PCL filament to a

level of over 20% in the final generations (Fig.

2c and Supplementary

Table 1). To better understand the effect of replication on the library

composition, we also modeled the number of correct sequence

reads per oligo using a negative binomial distribution. This analy

sis showed an escalation in the negative binomial over-dispersion

parameter from 1.0 in original library to 2.95 in F5 (Fig.

2d,e). This

increase in over-dispersion means that the oligos are becoming less

equally represented, which may be ascribed to the added thermal

stress during polymer extrusion/printing and increasing numbers

of PCR rounds

5,14

. In addition, we observed a significant increase

(Spearman ρ = 0.24, n = 600,000, P < 10

−9

) in the rate of substitution

errors as a function of the generation and a significant but modest

decrease in the rate of deletions and increase in the rate of insertions

(Spearman ρ

deletion

= −0.05, Spearman ρ

insertion

= 0.008, n = 600,000,

P < 10

−9

) (Supplementary Fig. 5). The length distribution of the

DNA oligos was quite similar and we did not observe a clear trend

(Supplementary Fig. 6). However, we did find a slight underrepre

sentation in later generations (Pearson r = −0.0375, n = 4.9 × 10

P < 10

−9

) of DNA oligos with higher numbers of guanine nucleo-

tides (Supplementary Fig. 7), which are known to be sensitive for

extrinsic stressors

. Despite these myriad imperfections, the DNA

Fountain strategy retrieved the file correctly for all generations,

showing the robustness of our strategy.

The economy of the DoT architecture is consistent with mass

production of goods with memory at negligible per-unit costs. Each

replication consumes only 0.3% from each bunny and yields suffi

cient DNA material to create 29 offspring bunnies. Therefore, even

if we restrict the number of replications to five generations, it is

theoretically possible to create at least (29/0.003)

= 8.44 × 10

bun-

nies without resynthesizing the DNA library. These results indicate

that despite the relatively high upfront costs to synthesize the file

for the first time ($2,500 in our case), the per-unit costs of synthesis

are likely to be negligible as mass production will require only one

synthesis event. To better understand the total costs per unit, we

conducted an analysis that also includes the costs of the other molec-

ular procedures, such as SPED preparation and PCR, using current

laboratory costs and an industrial scale (Supplementary Note 6).

Our results show that with the industrial scale costs, manufacturing

over 1,000 units of the same item, the cost of 1 MB DoT is smaller

than the cost of the polymer filament (Supplementary Fig. 8).

We envisage that the DoT architecture could be compatible with

a wide range of embedding materials. For such compatibility, the

DNA library should be miscible with the final material and sur

vive the melting temperature of the embedding material and the

material formation chemistry, which may include radical oxygen

species (RoS) or other types of aggressive chemical reactions (for

example, radical polymerization). Since DNA is not dissolvable in

most organic solvents, or polymeric systems, the encapsulation of

the DNA library into a dispersible carrier enables the mixing of the

DNA with such materials. As silica particles have a long use history

as a polymer filler

, the usage of SPEDs is advantageous

We sought to empirically test the robustness of our architecture

for various polymer preparation techniques. We first used quantita

tive PCR (qPCR) to measure the amount of DNA retrieved from the

SPED beads in PCL using various extrusion and 3D-printing tem-

peratures (Supplementary Fig. 9 and Supplementary Note 7). We

found that the amount of recovered DNA follows Arrhenius expres-

sion trends (Supplementary Note 8). For the extrusion process,

for each 30 °C increase in temperature, there is a drop of approxi-

mately one order of magnitude in the recovered DNA. On the other

hand, we observed only a relatively slow decline in the recovered

DNA with the increase of the 3D-printing temperature, which is

expected, as this process is much faster than extrusion and therefore

exposes the library to high heat only for a short amount of time.

The results also indicate that it is possible to use elevated tempera

tures for short cycle times during polymer processing, while suffer-

ing some losses of DNA

, which could be compensated by loading

more DNA into the polymer preforms. Such temperatures will allow

the use of DoT for selected polymer injection molding processes,

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

MPF1 F2 F3 F4 F5

Frequency

Perfect calls per oligo per million reads

Maximum restorable drop-out rate

Overdispersion: 1.04

Mean = 16.3

0.01 g 3.2 g

Parent

Identical Progeny

Drop-out rate

9 months

0510 15 20

0.05

0.10

0.15

0.20

Mean = 8.5

Overdispersion: 2.95

Frequency

020406080 100

0.01

0.02

0.03

0.04

0.05

0.06

Perfect calls per oligo per million reads

Mastrer library

Fig. 2 | DoT objects can be replicated for multiple generations. a, Just 10mg of the parent 3D-printed object was sufficient to recover enough DNA

with accurate building instructions (.stl file) to generate identical progeny. Scale bar, 1cm. b, The initial bunny and three ‘children’ printed with the DNA

instructions embedded in the parent DNA. The progeny also hold DNA instructions using the original DNA library. Scale bar, 1cm. c, For every new

generation, the file corruption rate increased, yet for the six experimentally implemented generations, it remained well below the file corruption rate

that the Fountain error correction code allowed. d, Oligo coverage frequency distribution for the master library before embedding the DNA and the fifth

descendant generation (F5). M, master library; P, parent.

NATURE BIOTECHNOLOGY | VOL 38 | JANUARY 2020 | 39–43 | www.nature.com/naturebiotechnology