Thermo‐Responsive Living Microspheroids Enable a Regenerative Living Disk–Drive System for DNA Data Storage

ABSTRACT

DNA offers exceptional information density and long‐term stability, yet its practical deployment is limited by destructive readout and the absence of a reusable, physically addressable architecture that connects nanoscale molecular information with macroscale device‐level data organization. Here, we present a regenerative Living Disk–Drive system based on thermo‐responsive engineered living memory microspheroids (ELMMs), in which data‐encoded bacteria are encapsulated as discrete, file‐level living storage units. Each ELMM contains a clonal bacterial population carrying both an information plasmid, which encodes 26 × 26 pixel icon payloads and one‐ to three‐color intracellular fluorescent retrieval indices, and a help plasmid that enables CRISPR–Cas12a/λ‐Red rewriting of the data sequence and retrieval tag. A lyophilized ELMM database forms the Living Disk, which is coupled to an Optical Retriever and desktop‐scale Living Drive for closed‐loop retrieval, regeneration, and database replenishment. Released bacteria regrow for downstream readout or rewriting, while a fraction is re‐encapsulated into new ELMMs. The tested system retains retrieval, regrowth, and sequence recovery after four months of ambient dry storage and 13 lyophilization–rehydration cycles. Model‐based performance estimates are reported only as theoretical architecture‐level bounds. These results establish an experimentally bounded yet extensible architecture for physically manageable and regenerative DNA memory.

A regenerative Living Disk–Drive system couples a lyophilized Living Disk with a fluorescence‐assisted Optical Retriever and desktop Living Drive to establish a closed‐loop DNA memory architecture. DNA‐encoded living file units (ELMMs) are optically addressed, thermally released, biologically regrown, and microfluidically re‐encapsulated, enabling retrieval–regeneration–replenishment cycles and programmable information updating within a physically manageable storage workflow.

1. Introduction

Owing to its exceptionally high information density and long‐term chemical stability, deoxyribonucleic acid (DNA) is widely regarded as a molecular data carrier with the potential to overcome key limitations of conventional silicon‐based storage [1, 2, 3, 4, 5]. Over the past decades, advances in encoding strategies [6, 7, 8], DNA synthesis [9, 10, 11, 12, 13, 14, 15], sequencing readout [16, 17, 18, 19], and molecular encapsulation [20, 21, 22, 23, 24, 25, 26, 27, 28] have collectively established the feasibility of DNA as an information medium [12]. However, despite the maturation of individual technologies, DNA storage largely remains at the laboratory‐scale proof‐of‐concept stage, with a substantial gap between experimental demonstrations and deployable, reusable, long‐term information systems.

This gap arises not from intrinsic physical or chemical limitations of DNA molecules, but from the lack of a reusable, physically addressable architecture capable of connecting nanoscale molecular data with scalable, macro‐level system operation [5, 29]. Most existing studies focus on the DNA molecule as the primary unit of manipulation, lacking standardized minimal storage units and corresponding database‐level operational workflows. As a result, it remains difficult to simultaneously achieve stable preservation, non‐destructive random access, and repeated reuse within a single architecture. In vitro DNA storage typically relies on polymeric [24] or inorganic [21, 22, 23] encapsulation to enhance molecular stability; however, this often comes at the cost of storage density, accessibility, and rewritability, making such systems primarily suitable for cold‐data archiving [30]. In contrast, in vivo DNA storage exploits cellular homeostasis and endogenous replication mechanisms, providing intrinsic advantages for information maintenance [27, 28, 31] and editing [32]. Beyond cold‐data archiving, in vivo systems are conceptually compatible with warm‐data applications involving low‐frequency access [33] and modification [32, 34]. Nevertheless, as database scale increases, in vivo systems continue to fall short in random access, file‐level management, and systematic operation.

Random‐access capability is widely regarded as a critical bottleneck limiting the practical deployment of DNA storage [35]. Mainstream sequence‐based retrieval methods, including polymerase chain reaction (PCR) amplification and sequencing [32, 36], rely on inherently destructive readout processes that consume the original DNA library [37]. Even with compensatory amplification, the inevitable introduction of bias limits the reliability of repeated access [38, 39]. Although solid‐phase PCR [24, 40, 41, 42] partially mitigates the loss of DNA, limited modifiability and rewriteability hinder dynamic information management. Meanwhile, gene‐editing technologies such as clustered regularly interspaced short palindromic repeats (CRISPR) enable sequence‐specific recognition [43, 44, 45], but they lack integrated, high‐efficiency physical separation and database‐level operational modules, and recent efforts have increasingly focused on this objective [46]. Recently, physical‐separation‐based retrieval strategies [38, 47], especially encapsulating DNA [38, 48, 49] or cells containing DNA [50] into discrete, tagged units and using flow sorting for file‐level access, have emerged as a promising approach for efficient, unified file retrieval. However, these strategies typically function as standalone modules rather than integrated components of a system‐level workflow, relying on manual intervention instead of seamless automation.

To transform DNA storage into a viable platform, a system must simultaneously meet four engineering requirements: standardized, operable and manageable minimal storage units [21, 22, 23, 24, 38, 48, 49, 50, 51]; a high‐density, low‐energy consumption database architecture [52, 53, 54]; a workflow for retrieval [49, 55], reuse [48], non‐destructive data extraction [55], DNA encapsulation and decapsulation [54], and rewriting [34]; and a compact, user‐operable hardware terminal [5, 56]. To date, no DNA storage system has simultaneously integrated these four requirements into a unified platform that connects standardized storage units, a low‐energy database format with potential for high‐density organization, retrieval, and information‐rewriting workflows, and compact, user‐operable hardware. This gap highlights the challenge of realizing a more operational DNA data‐storage architecture.

In our previous work, we established first‐generation engineered living memory microspheroids (ELMMs) as static, fluorescence‐addressable living file units fabricated by droplet microfluidics and retrieved by fluorescence‐assisted sorting [50]. That study demonstrated that DNA‐encoded living bacteria could be packaged into discrete microspheroids, preserved by lyophilization, and physically retrieved from mixed populations according to fluorescent labels. However, in that design, ELMMs primarily functioned as protected and sortable storage units that could be preserved, identified, and retrieved, rather than as regenerative system components that could be processed after retrieval and returned to a usable database. Several limitations prevented this first‐generation platform from functioning as a disk–drive‐like DNA storage system. First, the matrix was mainly designed for bacterial encapsulation and protection, rather than for mild, controllable release compatible with downstream operation. Second, the bacteria relied on a single information‐plasmid architecture in which the data sequence and fluorescent tag were co‐located, without an independent orthogonal editing module for programmable information erasure and replacement within the same living storage framework. Third, although ELMMs were defined as file‐level storage units, the previous work did not establish how ELMMs could be organized into a disk‐like database or managed, accessed, and replenished at the system level. Fourth, fabrication, sorting, and downstream handling remained largely manual, without an integrated device connecting post‐retrieval release, bacterial culture, re‐encapsulation, and database replenishment. Thus, the first‐generation ELMM platform answered whether fluorescence‐addressable living DNA file units could be constructed, but left open the system‐level question of whether such units could be organized, processed, regenerated, and replenished within a disk–drive‐like DNA information architecture.

Here, we address this question by developing a regenerative Living Disk–Drive system enabled by thermo‐responsive ELMMs, advancing DNA data storage from static file‐unit engineering toward a physically operable system architecture. Within this framework, data are no longer handled as anonymous molecular pools, but are compartmentalized into discrete, file‐level living storage units. Each ELMM contains an engineered bacterial population carrying the same encoded file and programmable fluorescent retrieval tags, thereby establishing a deterministic mapping between nanoscale molecular information and a physically addressable unit identity. This design allows file‐level target selection and physical retrieval to be guided by fluorescent signals, rather than by sequencing‐ or PCR‐based search of the entire data pool. Crucially, the thermo‐responsive gelatin matrix acts as a programmable physical barrier: it maintains spatial isolation during low‐temperature storage and undergoes a gel–sol transition upon mild heating, releasing the encapsulated bacteria for post‐retrieval regrowth and information‐carrier expansion. We define the recovery of data‐bearing biological information carriers through post‐retrieval bacterial regrowth, without de novo DNA synthesis or in vitro molecular amplification, as information regeneration. After release and regrowth, a fraction of the expanded bacterial population can be used for downstream readout or CRISPR‐mediated information rewriting, while the remainder can be re‐encapsulated into new ELMMs via integrated microfluidics to replenish the original database. This framework integrates DNA data preservation, fluorescence‐assisted physical retrieval, information regeneration, unit re‐encapsulation, database replenishment, and a programmable information‐updating interface within a single Living Disk–Drive architecture (Figure 1a).

FIGURE 1.

Workflow of engineered living memory microspheroids (ELMMs) and the Living Disk–Drive system. a, File‐level operation of a single living storage unit (ELMM), including retrieval, release, information regeneration, and rewriting. (i) Encoding and plasmid construction. Digital files are encoded into DNA sequences and assembled into an information plasmid together with fluorescent reporter cassettes serving as physical retrieval tags. A help plasmid enabling programmable information replacement is co‐transformed into bacteria, which are then encapsulated in a thermo‐responsive gelatin matrix to form ELMMs. (ii) Release of bacteria. Upon mild heating, the thermo‐responsive matrix undergoes a gel–sol transition and releases the encapsulated bacteria. (iii) Bacterial regrowth. Bacterial regrowth expands the data‐bearing biological information‐carrier population. For information regeneration and database replenishment, (iv.1) Re‐encapsulation. A fraction of the expanded bacterial population is re‐encapsulated by droplet microfluidics and low‐temperature gelation to generate refreshed ELMMs carrying the same stored file. For information replacement, (iv.2) Single‐round replacement of stored information and fluorescent retrieval tags. New DNA segments encoding updated data and corresponding fluorescent reporter cassettes are introduced. Cas12a, expressed from the help plasmid, cleaves the pre‐existing data and reporter region, while λ‐Red‐mediated homologous recombination installs the new sequence, enabling file replacement (for example, File 1–GFP to File 2–mCherry). (v.2) Re‐encapsulation of updated carriers. Re‐encapsulation yields new ELMMs carrying the updated information. b, Architecture and operation of the Living Disk–Drive system. The system consists of three functional components: ① Living Disk, a lyophilized ELMM database for ambient dry storage and portability; ② Optical Retriever, a fluorescence‐assisted physical retrieval module enabling file‐level selection and isolation of target ELMMs; and ③ Living Drive, an automated processing unit that supports bacterial release, regrowth, and re‐encapsulation, thereby enabling information regeneration, database replenishment, and a programmable information‐updating interface.

Multiple ELMMs carrying distinct files are organized into an ELMM database and converted into a Living Disk through lyophilization and low‐moisture, low‐oxygen nitrogen–vacuum packaging. This process exploits co‐vitrification of the gelatin matrix and cryoprotectants, enabling low‐energy dry preservation and portable transport under the tested storage conditions. The Living Disk can be stored as a standalone lyophilized medium or integrated with a microfluidic chip during lyophilization and packaging to form a chip–disk storage configuration (Figure 1b, ①). For operation, rehydration is initiated by adding culture medium to the chip region containing the Living Disk, after which the chip–disk assembly can be loaded into sorting equipment [57] (Figure 1b, ②). File‐level physical random access is achieved through fluorescence‐assisted sorting of target ELMMs based on predefined fluorescent indices, without requiring sequencing‐based search of the entire data pool. Retrieved ELMMs are then transferred to a desktop automated terminal, termed the Living Drive, which functions as an integrated biological processing module (Figure 1b, ③). Within this terminal, controlled heating induces ELMM dissociation and releases information‐carrying bacteria. Subsequent cellular proliferation expands the data‐bearing bacterial population, providing the biological basis for information‐carrier regeneration. A fraction of the expanded bacterial population can be used for downstream readout or rewriting, while the remainder can be re‐encapsulated into ELMMs to replenish the database. The release, culture, and droplet‐based re‐encapsulation processes are executed within the Living Drive. Together, these operations establish a closed‐loop retrieval–regeneration–replenishment workflow with a programmable rewriting interface, thereby transforming DNA data storage from a static molecular archive into a physically manageable and regenerative system‐level information platform.

The Living Disk–Drive system provides an experimentally bounded yet extensible architecture for regenerative DNA data storage, combining the Living Disk as a physically addressable lyophilized ELMM database with fluorescence‐assisted retrieval and downstream biological processing through the Living Drive. In the present proof‐of‐concept implementation, the system enables file‐level physical retrieval of 26 × 26 pixel icon payloads indexed by one‐ to three‐color intracellular fluorescent tags, post‐retrieval bacterial regrowth, re‐encapsulation of regenerated information carriers, and database replenishment. Based on the experimentally validated icon payload and a conservative sorting‐output rate, the current small‐file sorting‐stage retrieval rate is estimated to be ∼4.23 kB·s^{− 1}. Under explicit architecture‐level assumptions, including larger biological payloads and a theoretical target of 10⁴ indexed ELMMs, model‐based calculations further give a projected retrieval throughput of ∼1.19 MB·s^{− 1}, a logical disk capacity of ∼0.239 GB, and a corrected dried‐particle volumetric density of ∼2.9 × 10^{1 8} bytes·m^{− 3}. These architecture‐level values should be interpreted as model‐based estimates derived from component‐level assumptions supported by currently available technologies. Under the tested conditions, the system retained retrieval, regrowth, fluorescence‐index detectability, and sequence recovery after four months of ambient dry storage and after 13 lyophilization–rehydration cycles. Together, these results support the feasibility of organizing living DNA‐encoded file units into a physically manageable, regenerative disk–drive‐like system for DNA data storage.

2. Results

2.1. Plasmid‐Integrated DNA Storage System Enables Stable Regenerative and Rewritable In Vivo Data Memory

2.1.1. Construction, Stability, and In Vivo Rewriting of Plasmid‐Integrated DNA Storage System

To enable stable DNA‐encoded information propagation, fluorescence‐assisted retrieval, and programmable rewriting in living cells, we constructed an orthogonal dual‐plasmid system based on an Escherichia coli chassis (Figure 2a). This system functionally decouples information storage from information editing at the genetic level, providing a controllable molecular foundation for physical retrieval and information updating.

FIGURE 2.

Dual‐plasmid architecture enables stable propagation and programmable replacement of DNA‐encoded information. (a) Schematic workflow for integrating digital information and paired fluorescent reporter segments into plasmids and constructing the dual‐plasmid rewriting system. DNA‐encoded information and corresponding fluorescent tags are assembled into an information plasmid, while a help plasmid provides inducible plasmid‐editing functions. (b–d) Flow cytometry analysis of transformed Escherichia coli showing stable propagation of DNA‐encoded information across serial passages in representative single‐color (b), dual‐color (c), and triple‐color (d) strains after sequence integration, as indicated by the proportions of mCherry‐, EGFP‐, and tagBFP‐positive cells. (e) The information plasmid pUC57‐info‐cat‐Kan‐sfGFP and help plasmid p46Cpf1‐OP2 are co‐introduced into E. coli MG1655, yielding strain MG1655C‐K‐sfG. The help plasmid expresses λ‐Red recombinase under a tetracycline‐inducible promoter (P_tet ) and Cas12a under an arabinose‐inducible promoter (P_araBAD ). (f) Mechanism of programmable information rewriting. Two crRNAs guide Cas12a to generate double‐stranded DNA breaks within the target region of the information plasmid. A donor DNA fragment carrying the replacement information sequence (info‐banana), a new fluorescent reporter (mScarlet3), a tetracycline‐resistance marker, and ∼500 bp homology arms is integrated via λ‐Red‐mediated homologous recombination, enabling precise replacement of the original data cassette. (g) Conversion of the original information plasmid into the rewritten plasmid pUC57‐info‐banana‐Tet‐mScarlet3 following editing. The resulting strain MG1655B‐T‐M expresses mScarlet3, providing a direct phenotypic readout of successful information rewriting. (h,i) Raw flow cytometry data from the MG1655C‐K‐sfG (before rewriting) (h) and MG1655B‐T‐M (after rewriting) (i), where the ‘animal’ tag is expressed by the green fluorescent protein (sfGFP); the post‐rewriting ‘yellow’ tag is expressed by the red fluorescent protein (mScarlet3). (j) Quantification of green (sfGFP‐positive) and red (mScarlet3‐positive) bacterial populations before and after rewriting (n = 3 independent experiments; mean ± S.E.M.). (k) Sanger sequencing of DNA segments before and after rewriting, confirming the original and rewritten information, respectively.

The system comprises two components. The first component is the information plasmid (Info plasmid), constructed on a pUC57 backbone and serving as a modular data carrier that integrates three functional units: (1) a data payload region (Info), consisting of DNA sequences encoded from digital files (for example, images); (2) a retrieval tag region (Color seq), an expression module encoding specific fluorescent proteins that serves as a visual physical index for rapid localization of data content; and (3) a selection marker region (Anti1 seq), containing an antibiotic resistance gene to maintain plasmid stability and enable positive selection during subsequent information rewriting. Notably, the Info plasmid architecture is content‐agnostic, rendering it compatible with DNA data from diverse encoding strategies and facilitating scalability across various data types. In the current proof‐of‐concept implementation, one‐ to three‐color intracellular fluorescent indices were experimentally validated. Based on the combinatorial model described in Section S01, this regime supports an architecture‐level addressability estimate on the order of 10⁴ logical file identities.

The second component is a help plasmid designed to provide the gene‐editing machinery required for in vivo information rewriting. Under the control of an inducible promoter, this plasmid expresses the CRISPR–Cas12a nuclease together with the λ‐Red recombinase system. Cas12a introduces targeted double‐strand breaks in the existing Info plasmid, whereas the λ‐Red system mediates homologous recombination with donor DNA, enabling in situ information replacement.

As a proof of concept, we selected five images with distinct semantic features (e.g., “yellow”, “animal”, “aquatic”) and converted them into 26 × 26 pixel binary matrices, which were subsequently encoded into nucleotide sequences (Table S1). To support file‐level random retrieval, we established a fluorescence–semantics mapping scheme in which each file is assigned a unique combination of fluorescent protein expression cassettes (Table S2). Using a standardized cloning workflow, the encoding sequences were integrated into corresponding Info plasmids (Figure 2a and Figure S1) and transformed into E. coli TOP10 to generate an engineered bacterial library carrying distinct files (see Section S02 and Figures S2 and S3).

In the process of information regeneration, long‐term genetic stability is a key requirement for in vivo DNA data storage. To assess information regeneration stability, we performed approximately 216 generations of serial passaging (every 12 h; doubling time ≈20 min) on engineered bacteria carrying Info plasmids to evaluate fidelity during exponential replication. Sanger sequencing revealed that sequences identical to the original design were recovered at all checkpoints (Figure S4). Even after 216 generations of amplification, sequencing chromatograms remained clear with no overlapping peaks (Figure S5), indicating consensus‐level retention of the original information at the population level without detectable dominant mutations.

To quantify mutations at single‐base resolution, we performed next‐generation sequencing (NGS) on five files after 216 generations. No large‐fragment deletions or non‐specific amplification were detected in the Info plasmids (Figure S6). NGS revealed single‐base substitutions as the dominant error type; among all sequenced reads, 21.5%–40.5% were completely error‐free, and the weighted average error rate across all files remained below 0.2% (Figures S7 and S8). Although individual cells accumulated background mutations, consensus‐level sequence recovery enabled accurate reconstruction of the encoded information at the population level.

In parallel, the fluorescence indices used as physical retrieval indices remained stable. Flow cytometry showed that mono‐, bi‐, and trichromatic labeling systems maintained stable positive‐cell fractions after 216 generations, confirming the genetic robustness of the fluorescence–file mapping (Figure 2b–d and Figures S9 and S10).

We next tested whether the orthogonal dual‐plasmid system could provide a proof‐of‐concept interface for DNA data rewriting. Using Cas12a‐guided cleavage and λ‐Red‐mediated homologous recombination, we achieved in situ erasure and rewriting of Info plasmids. As an example, a crRNA was designed to target the Info plasmid encoding the “Cat” image in strain MG1655‐Cat (green fluorescence sfGFP, Kanamycin resistance Kan) (Figure 2e). To update the information to a “Banana” image, we constructed a donor DNA containing a new data sequence, a red fluorescent protein mScarlet3 expression cassette, a new selection marker (tetracycline resistance, Tet), and ∼500 bp homology arms on each side.

Following the induction of Cas12a and λ‐Red expression, the donor DNA was transformed into competent cells (Figure 2f). Cas12a, guided by crRNA, generated DSBs on the old Info plasmid, erasing the original data, retrieval tag, and selection marker; subsequently, the λ‐Red system mediated homologous recombination repair with the donor DNA, achieving a one‐step replacement of the data, fluorescent‐index, and selection‐marker modules to generate the new strain MG1655‐Banana (Figure 2g and Figures S11 and S12). After further screening using antibiotic culture, the phenotypic shift after rewriting was distinct: fluorescence microscopy and flow cytometry (Figure 2h–j and Figure S13) showed the bacterial population successfully transitioned from green to red fluorescence, with the percentage of green cells decreasing from 96.7% to 0.004%, and red cells rising from 0.001% to 67.2%. Sanger sequencing confirmed that the rewritten Info plasmid sequence perfectly matched the target “Banana” data (Figure 2k and Figure S14; sequences in Tables S3 and S4).

In summary, the orthogonal dual‐plasmid system enables programmable in situ erasure and rewriting of DNA‐encoded information at the plasmid level. Upon expansion of the retrieved cell population, one fraction can be preserved or re‐encapsulated to maintain the original information, whereas another fraction can serve as a biological substrate for downstream rewriting. This architecture provides a genetic foundation for updatable in vivo DNA memory, while iterative rewrite cycling and long‐term rewrite endurance remain to be evaluated in dedicated future studies (see Section S11).

2.2. File Fabrication With Replication and High‐Fidelity Retrieval Capability

2.2.1. High‐Throughput Fabrication, Thermo‐Responsive Release, and File‐Level Retrieval of ELMMs

To convert planktonic engineered bacteria into encapsulated, retrievable, and easily handled standardized storage units while restricting environmental dispersal [58], we developed a microfluidic gelatin‐based system to generate engineered living memory microspheroids (ELMMs). This system combines physical encapsulation with material responsiveness to convert microscopic biological carriers into physically addressable microscale file units suitable for file‐level operations. ELMMs were generated using a flow‐focusing microfluidic device. An engineered bacterial suspension containing 7.5% gelatin served as the dispersed phase and was injected into a fluorinated oil phase (Novec 7500) containing 0.1% Pico‐Surf surfactant. Precise control of the flow rate ratio enabled stable on‐chip generation of highly uniform microdroplets. The droplets were rapidly solidified by ice‐bath‐induced gelatin sol–gel transition through physical crosslinking, yielding structurally stable ELMMs (Figure 3a).

FIGURE 3.

Fabrication, thermal responsiveness, and file‐level retrievability of engineered living memory microspheroids (ELMMs). (a) Schematic illustration of ELMM fabrication using droplet microfluidics, in which engineered bacteria carrying DNA‐encoded information are encapsulated within a gelatin‐based thermo‐responsive matrix. (b) Representative optical micrograph of ELMMs generated by droplet microfluidics. (c) Diameter distribution of ELMMs, demonstrating uniform size control during fabrication (n = 100 ELMMs; mean ± S.E.M.). (d) Temperature‐triggered melting behavior of gelatin‐based ELMMs and release of encapsulated bacteria upon transfer from 4 °C to 37 °C. (e–g) After ELMM melting at 37°C and bacterial release, cells are cultured continuously for 8 h (e), OD600 values are measured to monitor bacterial proliferation (n = 3; mean ± S.E.M.) (f), and bacterial growth kinetics are fitted using a logistic model (g). (h) Flow cytometry fluorescence profiles of representative single‐color (mCherry⁺, EGFP⁺) and dual‐color (mCherry⁺EGFP⁺) ELMMs. (i) Schematic of the fluorescence‐based physical sorting workflow for a mixed population of mCherry⁺ and EGFP⁺ ELMMs. (j,k) Images (j) and flow cytometry profiles (k) of ELMM mixtures prior to sorting. (l,m) Images (l) and flow cytometry profiles (m) of EGFP⁺ ELMMs after fluorescence‐based sorting. (n,o) Images (n) and flow cytometry profiles (o) of mCherry⁺ ELMMs after fluorescence‐based sorting. (p) Quantification of the proportions of red (mCherry⁺) and green (EGFP⁺) ELMMs before and after sorting, confirming high‐purity file‐level physical retrieval (sample size n = 3; data shown as mean ± S.E.M.).

Image analysis showed that microfluidic fabrication produced highly monodisperse ELMMs (Figure 3b and Figure S15). The average microspheroid diameter was 59.96 µm with a coefficient of variation of 4.06% (Figure 3c). This dimensional uniformity enables high‐precision sorting by flow cytometry. The fabrication process exhibited high throughput, generating approximately 10⁴ ELMMs within 5 s. With in situ temperature‐controlled crosslinking, the complete encapsulation workflow was completed within 5 min. Together, these results demonstrate engineering‐scale scalability, enabling rapid encapsulation of data.

A key design feature of ELMMs is the use of the thermo‐responsive properties of the gelatin matrix [59] to enable programmable release. Gelatin undergoes a gel–sol transition at 37°C, which coincides with the optimal growth temperature of E. coli. ELMMs were transferred from 4°C (gel state) to 37°C for microscopic observation. The gelatin matrix liquefied and dissociated within 7 min, resulting in efficient release of the encapsulated bacteria (Figure 3d and Figure S16a). Quantitative image analysis showed an ∼8‐fold increase in the Z‐axis projection area within 7 min (Figure S16b), confirming that ELMMs undergo rapid decapsulation and release bacteria.

Bacterial viability was assessed in a 300 µL 2YT + AmpR culture medium following release. OD600 measurements and bright‐field imaging over 8 h at 37°C (Figure 3e and Figure S17) showed rapid entry of released bacteria into exponential growth, with growth kinetics closely following a standard logistic model (Figure 3f,g).

This behavior imparts ELMMs with two complementary functional states: (1) Cold Locking: at low temperature, bacterial metabolism is suppressed, and cells are physically constrained by the gelatin matrix, preventing structural disruption and bacterial leakage under storage‐state conditions; (2) Heat Activation: upon exposure to 37°C, matrix constraints are relieved, enabling rapid information regeneration through exponential proliferation. This thermocontrolled mechanism makes ELMMs living storage units inherently compatible with a retrieval–regeneration closed loop.

We next assessed the impact of ELMM encapsulation on data retrieval accuracy. The performance of planktonic bacteria vs. encapsulated ELMM was systematically compared using fluorescence‐activated sorting (FAS) (gating logic, Figure S18). Planktonic bacterial populations exhibit considerable signal variability because of variations in cell cycle [60], plasmid copy number [61], and metabolic heterogeneity [62]. Even under stringent gating, single‐color planktonic strains contained 3.16%–8.14% false‐negative (fluorescence‐deficient) cells (Figure S19a).

In contrast, ELMM encapsulation significantly improved the gating of the file retrieval signal. This improvement arises from signal integration: each ELMM contains multiple clonally identical bacteria, and its overall fluorescence reflects the aggregate of all internal cells, thereby averaging out stochastic single‐cell fluctuations. Flow cytometry showed that ELMMs formed compact, well‐separated clusters within mixed libraries (Figure S19b). The proportion of double‐negative (false‐negative) events dropped to 0%–0.39%, and the identification of target positive populations was markedly improved (Figure S19c,d).

To simulate more complex retrieval scenarios, a mixed library of red/green color strains was constructed. In the planktonic mixture, the double‐negative (false‐negative) rate reached 5.97%–6.14% (Figure S19e). Concurrently, non‐specific bacterial adhesion generated 0.22%–0.23% double‐positive (false‐positive) signals, compromising retrieval specificity. In contrast, when identical strains were encapsulated as ELMMs, their fluorescent‐index regions were clearly distinguished in flow cytometry, resulting in markedly improved gating separation. The double‐negative rate decreased to 0.069%–0.14%, with no double‐positive adhesion signals detected (Figure S19f). In summary, encapsulation of microscopic carriers (planktonic bacteria) into structurally stable, signal‐integrated macroscopic ELMM file units effectively mitigates biological noise and adhesion, providing a foundation for high‐fidelity, scalable, file‐level physical random access.

2.3. Ambient Dry Preservation and Cyclic Reuse of Files via Lyophilization–Rehydration

2.3.1. Ambient Dry Storage, Rehydration, and Repeated Access of Lyophilized ELMMs

Achieving energy‐efficient, long‐term ambient storage is essential for translating DNA data storage from laboratory proof‐of‐concept to engineering applications [23, 63]. Although cryogenic freezing (−80°C or liquid nitrogen) is considered the “gold standard” for biological preservation, its high energy demand, infrastructure cost, and operational complexity limit feasibility for large‐scale, distributed data archiving [64]. Inspired by natural anhydrobiotes [65, 66], we developed a lyophilization strategy using a gelatin–trehalose composite for protection. When combined with vacuum packaging under low‐moisture, low‐oxygen inert conditions, this approach supports ambient dry storage and repeated access of ELMM databases under the tested conditions.

Specifically, ELMMs were immersed in 0.4 M trehalose and subjected to standard freeze‐drying. Dried ELMMs were vacuum‐light‐shielded‐packaged in a nitrogen glove box to maximize isolation from humidity and oxygen. In this system, trehalose and gelatin act synergistically: trehalose substitutes for water through hydrogen‐bond interactions during dehydration, stabilizing membranes and proteins, while gelatin vitrifies during lyophilization, restricting molecular motion and degradation [67]. Together, this gelatin–trehalose formulation provides passive physicochemical stabilization during dry storage, while functional data recovery remains dependent on post‐rehydration cell recoverability, regrowth, and sequence recovery [67, 68].

To evaluate performance under practical storage conditions, red fluorescent ELMMs carrying the “Banana” file were lyophilized, packaged, and stored at room temperature in a desiccator for four months prior to rehydration testing. Confocal images revealed that rehydrated ELMMs largely recovered their original shape (Figure 4a,b and Figure S20a). Average diameter rapidly recovered to 63.59 µm, close to pre‐lyophilization values, with CV maintained at 3.11% (Figure 4c). This reversible structural recovery ensures ELMMs maintain uniformity after rehydration, providing a stable basis for subsequent FAS‐based sorting and file retrieval.

FIGURE 4.

Lyophilization–rehydration preservation and cyclic reuse of ELMM file units. (a,b) Confocal microscopy images of representative ELMMs encoding the “Banana” file before freeze‐drying (a) and after rehydration (b), showing preservation of microspheroid integrity and fluorescence. (c) Diameter distributions of ELMMs before freeze‐drying, after freeze‐drying, and after rehydration, showing pronounced shrinkage after lyophilization and size recovery after rehydration (n = 100). (d) Flow‐cytometry analysis of fluorescence intensity in mCherry⁺ ELMMs before lyophilization and after lyophilization, four months of ambient dry storage, and rehydration, indicating retention of fluorescent retrieval‐tag detectability under the tested condition. (e) Schematic of fluorescence‐based physical sorting of mixed single‐color ELMM populations (mCherry⁺ and EGFP⁺) after freeze‐drying, 4‐month ambient dry storage, then rehydration. (f,g) Images (f) and corresponding raw flow cytometry data (g) of mixed ELMMs prior to sorting. (h,i) Images (h) and flow cytometry profiles (i) of sorted EGFP⁺ ELMMs. (j,k) Images (j) and flow cytometry profiles (k) of sorted mCherry⁺ ELMMs. (l) Quantification of red (mCherry⁺) and green (EGFP⁺) ELMM proportions before and after sorting, confirming file‐level physical retrieval after the tested ambient dry‐storage period (n = 3; data shown as mean ± S.E.M.). (m) Temperature‐triggered melting and bacterial release of rehydrated ELMMs upon transfer from 4 °C to 37 °C. (n) Proliferation of bacteria released from rehydrated ELMMs during 8 h of continuous culture, quantified by OD measurements (n = 3; data shown as mean ± S.E.M.). (o) Confocal images of ELMMs after repeated freeze‐drying–rehydration cycles (1, 5, 9, and 13 cycles). (p) Quantification of fluorescence intensity across multiple freeze‐drying–rehydration cycles, demonstrating sustained stability of retrieval signals (n = 3 independent experiments; statistical analyses shown). (q) Diameter statistics of ELMMs after repeated freeze‐drying–rehydration cycles (n = 100 ELMMs). (r) Circularity measurements of ELMMs after repeated preservation cycles, indicating retention of microspheroid morphology (n = 54; data shown as mean ± S.E.M.; analyzed using two‐sided Student's t‐test or one‐way ANOVA). (s) Raw Sanger sequencing traces from rehydrated ELMMs after plate culture, supporting correct sequence recovery after the tested ambient dry‐storage and repeated lyophilization–rehydration conditions.

Scanning electron microscopy (SEM) revealed structural changes in ELMMs during lyophilization (Figure S20b). After lyophilization, ELMMs showed pronounced particle‐level shrinkage while maintaining intact spherical contours without obvious rupture or collapse. The average diameter decreased from 61.49 to 16.14 µm, with the CV reduced from 3.50% to 1.90% (Figure 4c). Under the conservative particle‐ensemble model described in Section S10, this dried‐state compaction corresponds to an estimated ∼25‐fold particle‐level utilization factor. This value provides a physical basis for increased dried‐state packing of ELMM particles, but should not be interpreted as a directly achieved system‐ or device‐level storage‐density gain, because practical density is additionally constrained by formulation loss, particle packing, aggregation, retrieval accessibility, and chip‐handling volume.

We then evaluated how freeze‐drying and rehydration affect core ELMM storage functions. Flow cytometry analysis indicated that over 98% of rehydrated ELMMs retained fluorescence (Figure 4d and Figure S20c), indicating retention of fluorescent retrieval‐tag detectability after storage and rehydration. In mixed libraries, rehydrated mCherry⁺ and EGFP⁺ ELMM populations remained clearly distinguishable and efficiently sorted, with >99.5% accuracy (Figure 4e–l and Figure S20d), showing that fluorescence‐based retrieval remained functional under the tested lyophilization–storage–rehydration condition. Rehydrated ELMMs also retained thermo‐responsive release, dissociating within 7 min (Figure 4m and Figure S21) and supporting normal exponential bacterial proliferation, with growth kinetics fitting the logistic model (Figure 4n and Figure S22). Sequencing of recovered colonies confirmed correct recovery of the target DNA sequence in the tested samples (Figure S23). Together, these results support retention of key ELMM functions after lyophilization and ambient dry storage, including fluorescent‐index detectability, thermo‐responsive release, bacterial regrowth, and sequence recovery.

Unlike conventional “read‐once” DNA storage, ELMMs enable system‐level re‐access, analogous to conventional electronic hard drives. To mimic frequent‐access scenarios, the ELMM database underwent 13 consecutive lyophilization–rehydration cycles. Despite some ELMM fragmentation and bacterial leakage, most ELMMs retained structural integrity after multiple phase transitions (Figure 4o and Figure S24), with diameter and circularity largely unaffected (Figure 4q,r). Fluorescence intensity remained within the effective gating range, with >98.9% positive signals (Figure 4p and Figure S25), supporting fluorescent‐index distinguishability after 13 cycles under the tested condition. Colony culture and sequencing confirmed correct recovery of the stored information in the tested samples (Figure 4s and Figure S27).

Using fluorescence decay measured during the repeated lyophilization–rehydration assay, we fitted an exponential decay model to estimate the fluorescence‐detectability window of the retrieval signal (Section S06 and Figure S26). This model suggested a projected detectability threshold beyond 40 cycles under the fitting assumptions; however, this projection should be interpreted only as a fluorescence‐signal detectability estimate, not as a validated functional lifetime or a demonstrated number of reliable access cycles for the complete Living Disk–Drive workflow. Minor ELMM fragmentation and gradual fluorescence decay were observed during repeated lyophilization–rehydration cycles, suggesting opportunities to further optimize strain stress tolerance, matrix composition, and lyophilization protocols. Longer‐term archival claims, including century‐scale preservation, would require dedicated accelerated‐aging studies and single‐ELMM regeneration‐fidelity assays. Therefore, in this work, the more‐than‐four‐month ambient dry‐storage result, 13‐cycle reuse experiment, and fluorescence‐decay projection are interpreted as tested endpoint evidence and conservative model constraints, rather than as definitive lifetime predictions (discussion and outlook, Section S12).

2.4. Construction and Validation of the Living Drive

2.4.1. Engineering a Living Drive for Automated Information‐Carrier Regeneration and Re‐Encapsulation

To standardize and automate the “thermo‐responsive release–bacterial regrowth–microfluidic re‐encapsulation” workflow, we developed a compact, integrated terminal, the Living Drive (Figure 5a,b). Fabricated through monolithic 3D printing and modular assembly (Figures S28 and S29), the device integrates biological reaction and microfluidic fabrication units in a compact, modular architecture, serving as an automated processing platform for post‐retrieval expansion and ELMM re‐encapsulation.

FIGURE 5.

Design of the Living Drive device and validation of automated regeneration and re‐encapsulation cycles. (a) Exploded view of the Living Drive, a palm‐sized, integrated fluidic processing device engineered for automated regeneration of ELMM file units. The system comprises: ① Solidification channel, a fixed pipeline for low‐temperature crosslinking of ELMM precursors during microspheroid preparation; ② Assembly base, the core structural module for ELMM melting, bacterial release, and cultivation; ③ Base, a mounting platform housing a heating element and thermoelectric cooler. Two modular microfluidic chips: ④ Mixing microfluidic chip, which combines cultured bacteria with 15% high‐concentration gelatin to produce 7.5% gelatin bacterial aqueous ink; ⑤ Flow‐focusing microfluidic chip, which shears the aqueous ink into ELMM precursors using an oil phase. Thermal management system: ⑥ Heating pads, maintaining 37°C for modules ②, ④, and ⑤; ⑦ Copper heat spreader, for uniform temperature transfer from ⑧ thermoelectric cooler, with waste heat dissipated via ⑨ water‐cooling system. The assembly workflow is detailed in Fig S29a. (b) Photograph of the fully assembled Living Drive device. (c) Schematic of the automated regeneration workflow implemented by the Living Drive. Retrieved ELMMs are thermally melted to release encapsulated bacteria, which undergo exponential proliferation in the culture pool. Amplified bacteria are subsequently mixed on‐chip with gelatin matrix to form an aqueous ink, processed through droplet microfluidics, and rapidly crosslinked at low temperature to generate regenerated ELMMs. (d) Growth kinetics of bacteria released from varying initial numbers of ELMMs following melting. Calibration curves indicate the time required to reach the standardized re‐encapsulation concentration of 2 × 10⁸ cells mL^{− 1} for different starting ELMM inputs (n = 4 independent experiments), enabling adaptive, input‐dependent timing control of the regeneration process. (e) Representative optical images of ELMMs generated in the first and tenth regeneration cycles during continuous, on‐chip multi‐round operation, demonstrating consistent microspheroid morphology across iterations. (f) Diameter distributions of ELMMs across ten consecutive regeneration cycles, showing low variation and stable size control. (g) Mean fluorescence intensity of regenerated ELMMs over ten release–regenerate–re‐encapsulation cycles, indicating retention of fluorescence‐signal detectability under the tested conditions (data from three independent experiments; instrument batch medium used as control).

The Living Drive consists of three functional submodules (Figure 5a): (1) Biological Reaction and Fluidic Handling Module: This module features a multifunctional assembly base integrating a micro‐culture pool with slots for two interchangeable microfluidic chips. The Mixing Chip combines amplified bacterial culture with a 15% gelatin matrix at a 1:1 ratio to form an aqueous precursor for re‐encapsulation, while the Flow‐focusing Chip uses oil‐phase shearing to generate highly monodisperse ELMM droplets. (2) Thermal Management System: A zonal temperature control network maintains the culture pool, mixing region, and droplet generation zone at 37°C to sustain bacterial activity and matrix fluidity. Simultaneously, a Peltier cooler with a copper heat spreader and water‐cooling preserves a low temperature in the downstream solidification channel, enabling rapid in situ droplet crosslinking. (3) Automated Fluidic Drive System: External pumps coordinate all channels to automate sample transport between modules (fluidic path schematic in Section S07 and Figures S29–S31).

Using this architecture, the Living Drive implements a standardized regeneration workflow (Figure 5c). Target ELMMs sorted by FAS are directly loaded into the culture pool, whose walls are pre‐coated with PF127 to prevent non‐specific bacterial adhesion. At 37°C, the ELMM matrix liquefies, releasing seed bacteria that undergo exponential amplification in the pool. Once a predefined bacterial density is reached, the culture is pumped into the Mixing Chip, combined with pre‐heated gelatin to form a 7.5% aqueous “ink,” and subsequently processed through the Flow‐Focusing Chip into microdroplets. These droplets rapidly gel in the low‐temperature channel, completing information‐carrier expansion and re‐encapsulation into refreshed ELMMs (Movie S1).

All components in contact with biological fluids, including the assembly base and microfluidic chips, are modular and single‐use, minimizing the risk of cross‐contamination. This design enhances biosafety, reduces operational complexity, and allows future engineering optimization (Movie S2).

In practice, the number of sorted ELMM “seeds” varies from a few to tens of thousands. We developed an adaptive culture model correlating initial input with growth kinetics by introducing 10, 100, 1000, and 10 000 ELMMs into the culture pool and monitoring OD600 in real time (Figure 5d and Figure S32). To reach the re‐encapsulation threshold (>2 × 10⁸ CFU·mL⁻¹), 10 000 ELMMs required 8.1 h, whereas 10 ELMMs required 18.1 h. This calibration allows the Living Drive to determine culture duration based on sorted ELMMs, ensuring consistent bacterial payload across re‐encapsulation batches.

To assess the engineering stability and durability of the Living Drive, we conducted 10 consecutive automated release–regenerate–re‐encapsulation cycles. In each cycle, ∼100 ELMMs from the preceding round were used as seed inputs, cultured over 16 h, and subsequently re‐encapsulated. Morphological and fluorescence analyses revealed highly stable performance across iterations: ELMMs from the 10th round showed no significant morphological difference from the first round (Figure 5e and Figure S33), with diameter coefficients of variation consistently maintained at a low level (CV < 5%) (Figure 5f). Importantly, average fluorescence intensity remained stable across all 10 cycles (Figure 5g and Figure S34), indicating retention of fluorescence‐signal detectability during repeated release–regenerate–re‐encapsulation operation under the tested conditions.

In summary, the Living Drive converts traditionally manual biological operations into a programmable, reproducible workflow, demonstrating the feasibility of information regeneration and fluorescence‐signal‐preserving automated release–regenerate–re‐encapsulation workflow. In the future, leveraging the maturity of microfluidics‐based CRISPR‐mediated genome editing [69], plasmid extraction units [70], and desktop third‐generation sequencers [71], the Living Drive adopts a modular design to ensure system compatibility. Its modular layout may facilitate future integration of additional modules, such as bacterial transformation, plasmid extraction, sequencing, and rewriting validation, but these functions were not integrated or validated in the current prototype. As a hardware prototype, the Living Drive provides a modular platform for future development of more automated living DNA storage workflows.

2.5. Integrated Living Disk–Drive System for File Retrieval and Regeneration

2.5.1. System‐Level Validation of File‐Level Random Access and Regenerative Retrieval

To enable automated, desktop‐compatible DNA data management, we developed a Living Disk–Drive system prototype integrating information storage, physical retrieval, and biological regeneration (Figure 6a). The prototype consists of three functional modules: the Living Disk, a lyophilized ELMM database supported by a carrier chip; the Optical Retriever, a fluorescence‐assisted physical retrieval module; and the Living Drive, an automated downstream processing module for release, bacterial regrowth, and re‐encapsulation. This modular organization provides a disk–retriever–drive‐like operational framework for physically manageable DNA data storage.

FIGURE 6.

Integrated Living Disk–Drive system for file‐level random access and regenerative retrieval. (a) Schematic overview of the Living Disk–Drive workflow, integrating a lyophilized ELMM database (Living Disk), fluorescence‐assisted physical retrieval (Optical Retriever), and downstream information‐carrier regeneration and re‐encapsulation (Living Drive) into a closed‐loop retrieval–regeneration–replenishment framework. (b) Commercial microfluidic sorting chip (Namocell cartridge) used as a carrier interface for Living Disk validation experiments. (c) Formation of a Living Disk by lyophilizing a mixed ELMM database under low‐moisture and low‐oxygen conditions, enabling ambient dry storage under the tested condition. (d) Rehydrated Living Disk after on‐chip medium injection, ready for interfacing with the optical retrieval system. (e) System‐level operation of file retrieval and regeneration. Following chip loading, target ELMMs are physically identified and isolated by a desktop fluorescence‐assisted sorting device, and the retrieved ELMMs are transferred downstream to the Living Drive for bacterial regrowth and re‐encapsulation. (f) Schematic of file‐level physical random access demonstrated using a mixed database containing two files: an mCherry⁺ “Banana” file and an EGFP⁺ “Cat” file. Target files are addressed by fluorescence‐based Boolean gating and retrieved as discrete ELMM objects. (g) File composition before and after sorting, showing enrichment of target ELMMs during physical retrieval. (h) Representative bright‐field images of ELMMs before and after sorting, showing physical separation of distinct file populations. (i) Sanger sequencing validation of retrieved files. 50 colonies were randomly selected from sorting plates for each target file. Sequencing results support file‐specific retrieval and sequence recovery, with target file sequences indicated by solid lines and rare non‐target background sequences indicated by dashed lines. Grayscale intensity reflects the relative proportion of retrieved file types.

For prototype validation, we used a commercial microfluidic chip (Namocell cartridge) as the physical carrier of the Living Disk (Figure 6b and Figure S35). The sample reservoir was loaded with ELMMs carrying distinct data files to form a mixed ELMM database. After loading, the chip‐associated ELMM database was lyophilized and vacuum‐packaged under low‐oxygen (<1.5 ppm) and low‐moisture (<0.1 ppm) nitrogen conditions (Figure 6c and Figure S36). By leveraging the protective vitrification effect of trehalose, the system supported room‐temperature storage of ELMMs over the tested period and tolerated environmental temperature fluctuations within the tested range (20°C–30°C). This reduces dependence on cold‐chain infrastructure under the tested storage conditions, potentially reducing energy consumption and maintenance complexity.

Downstream, the system incorporates two critical functional modules. The first is the Optical Retriever, implemented via a desktop fluorescence‐activated sorting device (Namocell Pala), which enables physical addressing and random access of target files by detecting multi‐channel fluorescent tags on ELMMs. The second is the Living Drive, the previously described automated fluidic terminal, responsible for receiving sorted ELMMs and executing the standardized “thermal release–bacterial amplification–microfluidic re‐encapsulation” workflow. Importantly, this system represents more than a hardware assembly; it constitutes a redefinition of the DNA data lifecycle. By expanding bacteria released from retrieved ELMMs and re‐encapsulating them into new file units, the system bridges single‐ELMM physical addressing with downstream biomanufacturing.

Leveraging this hardware integration, the Living Disk–Drive system establishes a standardized data‐access mode. Culture medium is injected into the Living Disk to rehydrate lyophilized ELMMs in situ within approximately 10 min (Figure 6d). The chip is then docked to the Optical Retriever. For performance interpretation, we distinguish the experimentally relevant sorting‐output rate from instrument‐level enrichment or event‐processing rates. According to the updated analysis in Section S08, a conservative sorting‐output rate of 50 ELMMs·s^{− 1} gives an experimentally relevant sorting‐stage query rate of approximately 4.23 kB·s^{− 1} for the current 26 × 26 pixel icon files. Under the conservative architecture‐level payload model, the projected sorting‐stage retrieval throughput is approximately 1.19 MB·s^{− 1}. Higher rates based on 50 000 ELMMs·s^{− 1} should be interpreted only as enrichment‐mode or event‐processing upper‐bound estimates, not as experimentally achieved end‐to‐end retrieval bandwidth of the current prototype. Upon detecting a target file that meets preset Boolean gating conditions, the Optical Retriever isolates the corresponding ELMM for downstream transfer to the Living Drive (Figure 6e and Figure S37 and Movie S3).

We first assessed the system's physical‐level operational resolution (Figures S38 and S39). In single‐droplet sorting tests (Figure S39), single‐ELMM sorting was achieved in 91.7% of 12 independent experiments, supporting the feasibility of single‐ELMM physical isolation. These results support the feasibility of single‐ELMM physical manipulation, while the sorting assay should be interpreted as a precision test of physical isolation rather than a complete single‐ELMM regeneration‐fidelity assay.

To further validate retrieval performance in a two‐file mixed database, we constructed a mixed Living Disk containing two file types: “Banana” (mCherry‐labeled) and “Cat” (EGFP‐labeled) (Figure 6f and Figure S40). After four months of ambient dry storage, the disk was rehydrated and subjected to dual‐channel fluorescence retrieval (Figure 6g and Figure S41). Bright‐field and fluorescence imaging showed clear separation of the two sorted ELMM groups, while quantitative image analysis indicated sorting purities exceeding 99.2% for both groups (Figure 6h and Figure S42). These findings support retention of fluorescent‐index detectability and file‐level physical retrieval after the tested ambient dry‐storage period.

At the molecular level, sorted samples were subjected to plating culture (Figure S43a) followed by Sanger sequencing (Figure S43b). For each file‐specific population (mCherry⁺ and EGFP⁺), three independent sorting and culture replicates were performed. From each replicate, 50 colonies were randomly selected for sequencing (Figure 6i). Across the three replicates, the lowest observed accuracy was 49 correct sequences out of 50 colonies for both the mCherry⁺ population (Banana file) and the EGFP⁺ population (Cat file) (Figure S43c). Overall, genotypic accuracy across all replicates exceeded 99%, supporting file‐specific retrieval, post‐sorting regrowth, and correct sequence recovery under the tested workflow.

In summary, the integrated Living Disk–Drive prototype combines a lyophilized ELMM database, fluorescence‐assisted physical retrieval, and downstream information‐carrier regeneration through the Living Drive. The current two‐file retrieval experiment demonstrates system‐level feasibility of file‐level physical access, post‐retrieval regrowth, and sequence recovery under the tested storage and retrieval conditions. These results establish a prototype framework for physically manageable and regenerative DNA memory, while larger library sizes, end‐to‐end automation, long‐term lifetime prediction, and single‐ELMM regeneration‐fidelity stress testing remain important directions for future validation.

3. Discussion and Outlook

Here, we present and validate a desktop‐scale Living Disk–Drive system enabled by thermo‐responsive engineered living memory microspheroids (ELMMs), advancing DNA data storage from a predominantly molecule‐centric paradigm toward a physically operable, compartmentalized, and regenerative system architecture. By integrating fluorescence‐assisted physical random access, thermo‐responsive release, biological regrowth of information carriers, ELMM unit regeneration, database replenishment, and a programmable information‐rewriting interface within a unified workflow, the system establishes an experimentally bounded framework for regenerative DNA data storage. Spanning storage, retrieval, release, regrowth, re‐encapsulation, and proof‐of‐concept information replacement, this platform provides a system‐level blueprint for physically manageable, renewable, and programmable DNA‐based data storage.

Early DNA storage strategies, whether based on solution‐phase DNA [36], solid‐phase substrates [24, 41], or molecular encapsulation [38, 72], were largely built around a static view of DNA as a storage medium. Once information was written and archived, subsequent access relied on PCR‐based amplification followed by sequencing, progressively depleting the database or introducing compositional imbalances over repeated readout cycles. This inherent “read‐destruction” behavior has long represented a central obstacle to the practical engineering of DNA storage systems [36]. The Living Disk–Drive system addresses this limitation by establishing a closed‐loop, thermo‐responsive living microspheroids‐enabled read–regenerate paradigm. By compartmentalizing information‐bearing living chassis within a thermo‐responsive gelatin matrix, the system enables thermally triggered release of bacterial carriers after physical retrieval and supports their post‐retrieval proliferation under controlled culture conditions. This workflow provides a potential route to reduce library depletion and amplification‐associated compositional drift, although quantitative long‐cycle database‐composition stability remains to be evaluated.

Building on this regenerative framework, we further integrate CRISPR‐based editing to address the intrinsic “write‐once” limitation of most physically encapsulated DNA storage media [24] and to provide a proof‐of‐concept interface for programmable information updating. In the current implementation, this experiment should be interpreted as a single‐round replacement of the information‐plasmid cassette, including simultaneous updating of the encoded data sequence, fluorescent retrieval tag, and selection marker.

Another key system‐level advance of this work lies in the physicalization of data retrieval. Most existing DNA storage systems remain tightly coupled to biochemical kinetics during access [36, 45], with latency and throughput fundamentally limited by the timescales of PCR amplification or sequencing. By contrast, the Living Disk–Drive system implements a physical random‐access strategy based on living fluorescent indices, allowing retrieval tags to persist and propagate with cells. This design structurally decouples file‐level retrieval from sequence decoding. As a result, for the physical retrieval stage, throughput is no longer dictated by PCR or sequencing reaction kinetics, but is primarily constrained by the performance of sorting and detection hardware. This shift redefines DNA data access from chemically selective, molecule‐centric retrieval toward spatially and tag‐encoded physical addressing, providing an engineering pathway for future performance scaling through dedicated, parallelized, or array‐based retrieval hardware.

A systematic comparison with representative DNA storage systems is summarized (Table 1), with an emphasis on integrated system‐level capabilities—including storage‐unit organization, retrieval mode, preservation format, regeneration workflow, rewriting interface, and operational integration—rather than isolated theoretical performance metrics. In the present proof‐of‐concept implementation, the experimentally validated payload is a 26 × 26 pixel icon indexed by one‐ to three‐color intracellular fluorescent labels. Using this validated small‐file payload and the conservative sorting‐output rate of the commercial Namocell device, the experimentally relevant sorting‐stage query rate is estimated to be ∼4.23 kB·s^{− 1} (Section S08). Under a literature‐supported large‐payload assumption, architecture‐level modeling gives a projected sorting‐stage query rate of ∼1.19 MB·s^{− 1}, a logical disk capacity of ∼0.239 GB for 10⁴ indexed ELMMs, and a corrected dried‐particle volumetric density of ∼2.9 × 10^{1 8} bytes·m^{− 3} under stated assumptions (Sections S08 and S10). These values should be interpreted as model‐based estimates or architecture‐level upper‐bound projections, not as experimentally achieved end‐to‐end device performance. Experimentally, under the tested four‐month ambient dry‐storage and 13‐cycle lyophilization–rehydration conditions, the current system retained key functional endpoints, including retrieval, regrowth, fluorescence‐index detectability, and sequence recovery. Longer‐cycle projections based on fluorescence decay should be viewed as predictive estimates of retrieval‐signal detectability, rather than validated functional lifetime or rewrite‐endurance limits.

TABLE 1.

System‐level comparison of representative DNA data storage architectures. Metrics include storage carrier state, access paradigm, random‐access method, access speed, theoretical data density, preservation conditions, information regeneration, information rewriting, non‐destructive access, and engineering integration.

Work	Storage carrier	Encapsulation material & time	Access paradigm	Random access method	Access speed	Max retrievable files (theoretical)	Data density (theoretical)	Preservation	Metabolic activity during storage	Mutation risk	Information regeneration capability	Information rewriting	Non‐destructive access	Engineering integration level	Genetically modified microorganism (GMM) encapsulation
Organick, Ang 2018 [36] (in vitro)	DNA pool (free oligos)	None	Molecular pool	PCR‐based	Slow (∼20–30 thermal cycles)	35	∼102 5 bytes/m3	Dehydrated/frozen	N/A	Low (chemical degradation)	No (PCR alters original pool balance)	No	No	Low	N/A
Xu, Ma 2021 [111] (in vitro)	DNA array	None	Fixed spatial array	Location‐based	Slow	4 oligos	∼102 0 bytes/m3	Easily degraded	N/A	Low	No	No	No	Low	N/A
Grass, Heckel 2015 [72] (in vitro)	Silica‐encapsulated DNA	Silica (4 days)	Static archive	None	None	None	∼102 4 bytes/m3	RT, ∼2000 years	N/A	Low	No	No	No	Low	N/A
Banal, Shepherd 2021 [38] (in vitro)	Silica sphere + ssDNA label	Silica (4 days)	Molecular sorting	FAS‐based	∼750 MB/s (theoretical, high‐end FACS)	∼101 5	∼102 4 bytes/m3	Label instability	N/A	Low	No	No	No	Medium	N/A
Bögels, Nguyen 2023 [48] (in vitro)	Thermo‐responsive microcapsules	Polymer (2 days)	Molecular pool	Multiplex PCR	Slow (∼20–30 cycles)	13 million oligos	Unmeasured	Low temperature, ∼1 month	N/A	Low	No	No	No	Low–Medium	N/A
Mao, Wang 2023 [112] (in vitro)	MOF‐encapsulated DNA	MOFs (5 min)	Static archive	None	None	None	Unmeasured	RT, ∼10 years	N/A	Low	No	No	No	Low	N/A
Shipman, Nivala 2017 [31] (in vivo)	CRISPR spacer array	Cell membrane	Temporal biological recording	None	None	Unmeasured	Unmeasured	Subculture/−80°C	Active	High (evolutionary drift)	No (mutation accumulates during division)	No	No	Low	No
Chen, Han 2021 [28] (in vivo)	Artificial chromosome	Cell membrane	Genomic integration	None	None	Unmeasured	Unmeasured	Subculture/−80°C	Active	High	No	No	No	Low	No
Sun, Dong 2023 [33] (in vivo)	Genomic long DNA	Cell membrane	Genomic integration	None	None	Unmeasured	Unmeasured	Subculture/−80°C	Active	High	No	No	No	Low	No
Hou, Qiang 2024 [34] (in vivo)	Shuttle plasmid	Cell membrane	Molecular key–based access	gRNA‐based	∼4 days	∼105	∼102 1 bytes/m3	Subculture/−80°C	Active	High	Limited (via cultivation)	Yes	No	Medium	No
Li, Yang 2024 [5] (DNA‐DISK)	Enzymatically synthesized oligos	Agarose (on‐chip)	Microfluidic array	Spatial addressing	4.4 min/bit (latency)	High	Unmeasured	On‐chip solid state	N/A	Low	No	No	No	High	N/A
This work (Living Disk–Drive system)	ELMM array	Matrix material (5 min) with decapsulation time (7 min)	Physical spatial disk architecture Model‐based logical capacity: ∼0.239 GB per 104‐ELMM Living Disk	FAS‐based physical extraction	Model‐based sorting‐stage estimate: ∼1.19 MB·s− 1 at 50 ELMMs·s− 1; 50,000 events·s− 1 is enrichment‐mode upper‐bound only	Demonstrated 1–3 color labels; theoretical indexing target ∼104; mathematical upper bound ∼101 6	Corrected dried‐particle model: ∼2.9 × 101 8 bytes·m− 3; prototype apparent density lower	Lyophilized, RT (>4 months experimentally verified)	None (dormant)	Present but bounded; founder‐bottleneck and storage‐induced variants require single‐ELMM validation	Yes (read–regenerate closed loop)	Yes	Yes	High (system‐level)	Yes (designed to reduce GMM leakage under tested storage‐state conditions)

As a proof of concept, the current implementation of the Living Disk–Drive system exhibits several engineering limitations. At present, the integrated chip–disk storage format relies on commercial microfluidic sorting chips that were adopted for validation rather than optimized for high‐density Living Disk storage. As a result, system‐level storage density is constrained, and mechanical reliability may be affected during freeze‐drying and rehydration processes. In addition, the retrieval module depends on commercial sorting instrumentation operating under a limited sorting‐output configuration, leading to relatively modest sorting speeds. Although idealized enrichment‐mode or highly parallelized hardware scenarios may yield much higher model‐based sorting‐stage throughput, the current experimentally relevant sorting‐output estimate remains far below the gigabyte‐per‐second regime. Furthermore, while single‐round CRISPR–Cas12a/λ‐Red‐mediated information‐plasmid cassette replacement has been experimentally validated, repeated rewrite cycling and fully automated rewriting have not yet been incorporated into the closed‐loop Living Disk–Drive operation (Section S11 and Table S5). Importantly, these limitations reflect the early‐stage nature of the current hardware and materials implementation, rather than fundamental constraints imposed by the regenerative system architecture itself.

Another important reliability boundary is the founder‐bottleneck effect during regeneration from a single ELMM. The population‐level passaging experiments demonstrate that consensus information can be recovered during bulk bacterial proliferation, but they do not fully substitute for regeneration‐fidelity testing at the single‐ELMM level. Because each ELMM contains a limited number of clonal bacteria, with an average of approximately 11 cells in the current formulation, a rare mutation or storage‐induced lesion present in one founder cell could, in principle, become enriched during post‐retrieval release and amplification [73, 74]. The integrated retrieval experiments provide workflow‐level and colony‐resolved supportive evidence, with genotypic accuracy exceeding 99% across tested populations, but they should not be interpreted as a systematic single‐ELMM stress test. Future reliability assessment should therefore combine independent single‐ELMM sorting, separate release and regeneration, colony‐resolved sequencing or deep amplicon sequencing, and redundancy strategies such as multi‐ELMM file replicas, increased founder‐cell number, consensus recovery, and error‐correction coding [7, 8, 36, 72].

Meeting the demands of petabyte‐ to zettabyte‐scale data archiving will require future iterations of the Living Disk–Drive system to advance across multiple, tightly coupled dimensions, including storage‐unit design, indexing capacity, payload scalability, long‐term preservation, retrieval hardware, and system autonomy. At the level of individual storage units, future ELMMs may be engineered as smart living materials with multifunctional capabilities, enabling improved low‐energy preservation, controlled release, repeated operation, and integrated quality monitoring [75, 76].

First, the choice and design of the living chassis directly shape key system performance metrics, including retrieval efficiency, environmental robustness, genetic stability, and per‐unit storage capacity. In the current implementation, the experimentally validated fluorescent indexing regime is limited to one‐ to three‐color intracellular labels [50, 77, 78, 79]. Although modern spectral flow cytometry can provide high‐dimensional detection capacity [80, 81], this instrument‐level capability should not be equated with the stable intracellular co‐expression of a large number of fluorescent proteins in a single Escherichia coli cell. In bacterial hosts, higher‐order fluorescent‐protein multiplexing is constrained by host metabolic burden, plasmid‐maintenance burden, protein maturation efficiency, spectral overlap, signal‐deconvolution accuracy, and cell‐to‐cell expression heterogeneity [80, 81]. Therefore, we use 10⁴ independently indexed ELMMs as a conservative architecture‐level target for a logical Living Disk, rather than as a fully validated database scale in the present prototype. Further expansion of addressability will likely require not only improved spectral discrimination, but also orthogonal non‐biological indexing layers, such as microspheroid size [82], morphology [83], material‐composition barcodes [84, 85], embedded optical dyes [86], quantum‐dot signatures [87], or Raman/SERS‐active matrix labels [80, 81]. These physical indexing strategies could reduce reliance on intracellular fluorescent‐protein co‐expression while preserving the Key–Value Mapping logic of signal‐guided retrieval (Section S01).

Beyond indexing capacity, conventional chassis such as E. coli impose limitations on long‐term environmental robustness and large‐payload maintenance. Larger payloads may eventually be achieved through lower‐copy vectors, genomic integration, artificial chromosomes, alternative microbial chassis, or eukaryotic hosts [28, 88, 89]. For example, Saccharomyces cerevisiae and yeast artificial chromosomes provide a possible route for increasing per‐cell payload capacity [28]. However, such strategies should be interpreted as future development pathways rather than current ELMM performance, because they would require independent validation of DNA maintenance, fluorescence‐index stability, viable regeneration, physical retrieval, and repeated re‐encapsulation within the ELMM context. For storage over substantially longer timescales, alternative host organisms may become important future chassis candidates [88, 89]. Spore‐forming Bacillus species possess specialized dormant‐state protection mechanisms that are mechanistically distinct from the vegetative E. coli/gelatin/trehalose formulation used here [90, 91, 92, 93]. Although such organisms hold promise for inspiring future designs for long‐term, “living storage,” incorporating this principle into a Living Disk–Drive system requires further design and investigation.

Second, equally important for system performance and scalability is the design of the ELMM matrix material and its structural organization [58, 94, 95]. Future generations of ELMMs need not be limited to the current single thermo‐responsive triggering scheme. For instance, photo‐cleavable moieties, such as o‐nitrobenzyl groups, could be incorporated into the hydrogel network to enable wavelength‐specific cleavage and release [96], thereby improving spatiotemporal control during retrieval operations. In parallel, the integration of nanoprobes that are sensitive to environmental cues—such as humidity or oxidative stress [97]—could introduce intrinsic self‐reporting capabilities. In this scenario, ELMMs could exhibit optical or colorimetric changes in response to environmental perturbations, providing a direct readout of database integrity without destructive sampling. To further balance biosafety with long‐term system stability, a core–shell ELMM architecture may be advantageous [58]. In such a design, a dense protective shell, for example based on polydopamine or silica, could provide sustained physical containment, while the core would house living bacteria responsible for proliferation and information processing [98]. During retrieval, the shell could be selectively removed or permeabilized under defined stimuli, releasing the core for amplification and readout.

Third, although this study demonstrates that the gelatin–trehalose formulation preserves ELMM functionality through water replacement, vitrification, and reduced molecular mobility, longer‐term preservation and repeated reuse will require coordinated optimization of the bacterial chassis, ELMM matrix, microspheroid architecture, cryoprotectant formulation, packaging conditions, and lyophilization–rehydration protocol [35, 88, 89, 99, 100, 101]. The current E. coli/gelatin/trehalose system should be understood primarily as a passive physicochemical stabilization strategy during dry storage, rather than as a system relying on active DNA repair during freeze‐dried dormancy. Any cellular repair activity, if it occurs, is expected to take place after rehydration in recoverable cells, rather than during the dry storage state. Systematic studies on long‐term preservation of engineered living materials remain limited [102], particularly with respect to storage‐associated degradation pathways, cycle‐associated damage accumulation, and the molecular mechanisms governing repeated lyophilization–rehydration processes [103]. In the present system, the matrix material, microbial chassis, ELMM structure, packaging format, and Living Disk configuration have not yet been optimized for maximum storage lifetime, repeated access, or regeneration fidelity. Therefore, the current four‐month ambient dry‐storage result, 13‐cycle lyophilization–rehydration assay, and fluorescence‐decay projection should be interpreted as tested endpoint evidence and model constraints, rather than as a validated lifetime prediction. Future work should combine accelerated‐aging experiments, glass‐transition‐temperature and water‐activity analyses, residual moisture and oxygen control, and regeneration‐fidelity assays at both the single‐ELMM and whole‐Living‐Disk levels to establish a quantitative lifespan‐assessment framework for functional ELMM‐based storage units. To support this direction, we have added a preliminary preservation framework (Section S12), which defines key functional endpoints and provides a reference structure for future systematic evaluation. Once sufficiently large experimental datasets become available, data‐driven or machine‐learning‐assisted approaches may further help screen and optimize multidimensional design parameters, including microbial strains, matrix materials, cryoprotectant formulations, packaging conditions, and lyophilization process parameters [104, 105, 106]. Together, these efforts would support a standardized quality‐control framework spanning ELMM fabrication, lyophilization, dry storage, rehydration, regrowth, sequence recovery, re‐encapsulation, and database replenishment.

At the system level, further expansion of Living Disk–Drive functionality will require coordinated evolution across the macroscopic disk architecture, retrieval modules, and downstream Living Drive platform. A particularly important development direction is the redesign of the microfluidic chip hosting the Living Disk. In the present prototype, the commercial sorting chip primarily functions as a handling and retrieval interface, rather than as a storage‐density‐optimized disk medium. Future dedicated Living Disk chips should reduce rehydrated handling volume, reservoir dead volume, tubing loss, aggregation, and retrieval‐accessibility penalties, thereby narrowing the gap between the corrected dried‐particle architecture‐level density and the apparent operational density of the current chip‐interfaced prototype. Under the model assumptions (Section S10), a 10⁴‐ELMM logical Living Disk corresponds to a sub‐mm³ dried‐particle volume, but the actual device‐level density will depend on packing, loading, rehydration, and retrieval‐interface design. By spatially organizing multiple Living Disks, a data‐center‐scale DNA storage architecture could be envisioned, in which individual disks are first selected by spatial location using microfluidic routing [52, 53] and subsequently subjected to FAS‐based file‐specific retrieval within each selected disk. Alternatively, optimization of chip materials and bonding processes to improve low‐temperature fracture resistance, together with standardized rehydration and loading protocols, could enable a plug‐and‐play, portable chip–disk storage configuration.

In parallel, advancing retrieval modules toward array‐based architectures represents a critical pathway for increasing system throughput. Multi‐channel retrieval platforms based on acoustofluidics [107], dielectrophoresis [108, 109], or optical array technologies [110] could reduce system footprint and cost while improving sorting‐stage throughput. However, any transition from the current kB–MB·s^{− 1} model‐based regime toward substantially higher bandwidths will require dedicated hardware validation, parallel retrieval architectures, and end‐to‐end accounting of recovery, regrowth, sequencing, decoding, and quality‐control bottlenecks.

Furthermore, integration of the Living Drive system with lab‐on‐a‐chip technologies offers a path toward more automated operations, including unattended amplification, re‐encapsulation, quality control, and, in future versions, information rewriting. While the current implementation employs a modular design—using two replaceable microfluidic chips for aqueous‐ink preparation and ELMM generation to validate modularized production—the information rewriting step remains offline, relying on electroporation and cell recovery. Future iterations may integrate plasmid extraction, CRISPR reagent delivery, cellular transformation, sequencing, and validation modules directly into replaceable microfluidic chips [69]. These modules could interface with the main unit and couple to benchtop third‐generation sequencing platforms [52], establishing a more complete automated loop spanning information updating, validation, regeneration, and replenishment.

With continued evolution of the Living Disk architecture, the systematic development of dedicated sorting technologies, deeper integration of the Living Drive with lab‐on‐a‐chip platforms, and the incorporation of computer vision and AI‐driven feedback control, the Living Disk–Drive system could mature into a more integrated living bio–hard‐drive platform capable of increasingly automated biological data operations.

The Living Disk–Drive is not intended to replace high‐performance silicon‐based storage, but rather to complement it by offering a distinct set of attributes, including low‐energy dry preservation, file‐level physical random access, post‐retrieval biological regeneration, and database replenishment, particularly for cold and low‐frequency warm data tiers. Relative to conventional in vitro DNA storage, the system provides structural advantages in information management and operational sustainability, including dormancy‐enabled preservation, resuscitation‐driven data‐carrier expansion, in vivo CRISPR‐compatible information modification, unit‐level encapsulation and decapsulation, and automated workflows that reduce manual intervention.

In application domains such as data‐center cold storage, deployment in extreme environments, and long‐term knowledge archiving, properties such as zero‐power retention during tested dry‐storage intervals, ambient‐temperature stability under defined conditions, and potentially repeatable non‐destructive retrieval–regeneration workflows may translate into favorable lifecycle energy consumption and reduced maintenance burden after further scale‐up validation. More broadly, this work outlines an engineering pathway for treating biological systems as programmable information infrastructure. By enabling data‐bearing biological carriers to be regenerated, physically reconstituted, and programmably updated within living substrates, the Living Disk–Drive introduces a regenerative paradigm for constructing sustainable DNA data‐storage ecosystems.

4. Methods

4.1. System Architecture and Operational Paradigm of the Living Disk–Drive System

A Living Disk–Drive system was constructed to enable in vivo DNA data storage, file‐level physical random access, non‐destructive repeated access, biological information‐carrier regeneration, unit re‐encapsulation, and database replenishment within a unified engineering framework. The system employs engineered living memory microspheroids (ELMMs) as the minimal physical storage units, organizes ELMMs into Living Disks as tangible biological storage pools, and performs file‐level release, bacterial amplification, and re‐encapsulation through a downstream Living Drive module. Programmable information updating was enabled by the orthogonal dual‐plasmid system and experimentally validated as a single‐round information‐plasmid cassette‐replacement process, rather than as a fully automated multi‐cycle rewriting function of the current Living Drive prototype.

The system comprises one fundamental storage unit and three functional modules: (i) ELMMs serving as the minimal file units; (ii) Living Disks functioning as lyophilized dry‐storage assemblies; (iii) a flow‐based analytical platform enabling file‐level retrieval based on fluorescent physical tags; and (iv) a Living Drive module responsible for regeneration and database replenishment. Unless otherwise specified, all storage, retrieval, amplification, replenishment, and statistical analyses were performed with a single ELMM as the smallest operational and quantitative unit.

Within this framework, each ELMM was defined as the physical carrier of a single digital file. Each ELMM encapsulated a population of engineered Escherichia coli cells harboring identical information plasmids and constitutively expressing fluorescent proteins, thereby forming stable physical index tags recognizable by optical systems. ELMMs corresponding to different files were physically mixed in a disordered manner to constitute a Living Disk. Living Disks were lyophilized as intact entities for long‐term storage and, upon rehydration, loaded onto microfluidic chips to enable file‐level physical random access.

As the downstream functional module, the Living Drive converted retrieved ELMMs into expandable living replicas and regenerated new ELMMs through automated microfluidic workflows. This workflow supported biological expansion of data‐bearing carriers and database replenishment, establishing a closed‐loop operational paradigm of retrieval, regeneration, and replenishment.

4.2. Construction of Living Information Carriers: A Dual‐Plasmid Storage–Rewriting System

4.2.1. Overall Design of the Dual‐Plasmid System

An orthogonal dual‐plasmid system was constructed in Escherichia coli (MG1655 or DH5α) to enable stable DNA information storage, physical indexing, and in situ rewriting. The system consisted of an information plasmid and a help plasmid.

The information plasmid was based on the high‐copy‐number vector pUC57 and carried encoded artificial DNA data sequences. A fluorescent protein gene was constitutively expressed on the same plasmid to serve as a file‐level physical index tag. Each fluorescent label was uniquely mapped to a corresponding data file, enabling physical random access. A selectable antibiotic‐resistance marker was included in each information plasmid to support plasmid maintenance and positive selection during information replacement, with the specific marker depending on the construct.

The help plasmid p46Cpf1‐OP2 (Hangzhou Suttone Biotechnology Co., Ltd.) expressed Cas12a under L‐arabinose induction and the λ‐Red recombination system under tetracycline induction. Under non‐induced conditions, the help plasmid did not interfere with information plasmid replication. Upon induction, site‐specific cleavage and homologous recombination enabled single‐round replacement of the information‐plasmid cassette in the tested workflow.

All plasmid constructs were verified by restriction digestion and Sanger sequencing (Sangon Biotech, Shanghai, China). Complete sequences of rewriting‐related plasmids are provided in the Supporting Information, and all information plasmids have been deposited on Addgene.

4.2.2. Construction of Donor DNA

Donor DNA fragments used for information rewriting (up‐info‐banana‐Tet‐mScarlet3‐down) were constructed by in vitro overlap PCR. The fragment sequentially consisted of an upstream homologous arm, the information sequence banana, the red fluorescent protein mScarlet3 (which exhibits higher brightness and photostability than mCherry), a tetracycline resistance gene (Tet), and a downstream homologous arm, with a total length of 3658 bp. After construction, the fragment was verified by Sanger sequencing using primers targeting both ends (constructed and validated by Hangzhou Suttone Biotechnology Co., Ltd.).

4.3. Construction of Single‐ and Dual‐Plasmid Bacterial Strains

4.3.1. Construction of Single‐Plasmid Strains (Heat‐Shock Transformation)

Information plasmids were transformed into Escherichia coli DH5α using the CaCl₂‐mediated heat‐shock method. Bacterial cultures were grown in LB medium at 37°C with shaking at 220 rpm to an OD₆₀₀ of approximately 0.5, incubated on ice for 10 min, and harvested by centrifugation at 5000 × g for 10 min at 4°C. Cell pellets were resuspended in ice‐cold 0.1 M CaCl₂ and incubated on ice for 30 min to prepare competent cells.

For transformation, plasmid DNA (2 ng·µL^{− 1}) was mixed with competent cells and incubated on ice for 30 min, followed by heat shock at 42°C for 45 s and immediate cooling on ice for 2 min. SOC medium (1 mL, pre‐warmed) was added, and cells were allowed to recover at 37°C with shaking at 220 rpm for 1 h. Recovered cells were plated onto LB agar plates containing ampicillin (100 µg·mL^{− 1}) and incubated at 37°C for 16 h. Single colonies were selected for sequencing verification.

4.3.2. Construction of Dual‐Plasmid Strains (Electroporation)

Helper plasmids were introduced into previously constructed information‐plasmid strains by electroporation.

Information‐plasmid strains were cultured at 37°C for 1 h, plated onto 1.5% agarose plates containing kanamycin (50 µg·mL^{− 1}), and incubated at 37°C for 12 h. Colonies were subsequently inoculated into LB medium containing kanamycin and grown overnight at 30°C with shaking at 220 rpm.

Cells were harvested by centrifugation at 4°C and 4000 × g for 2 min, washed twice with ice‐cold 10% glycerol, and finally resuspended in 200 µL ice‐cold 10% glycerol. Aliquots of 40 µL were prepared, and 100 ng of help plasmid DNA was added to each aliquot. After incubation on ice for 1 min, the mixtures were transferred to 2‐mm‐gap electroporation cuvettes (Bio‐Rad, USA) and electroporated using a Bio‐Rad Gene Pulser Xcell (2.5 kV, single pulse).

Following electroporation, cells were recovered at 30°C in medium containing kanamycin (50 µg·mL^{− 1}) and chloramphenicol (12.5 µg·mL^{− 1}) and plated onto LB agar plates with dual antibiotics. Plates were incubated at 30°C for 24 h. Transformants were screened by PCR and plasmid extraction to confirm coexistence of both plasmids.

4.4. Digital Information Encoding and “Fluorescence–Semantic” Mapping

Digital files (represented primarily by image data) were first converted into fixed‐size binary matrices and compressed using run‐length encoding to reduce redundancy. The compressed bitstreams were then mapped to quaternary nucleotide sequences, generating DNA data payloads suitable for chemical synthesis and biological replication.

During encoding, pseudo‐random encryption seeds were introduced, and each file was assigned a unique primer identifier for downstream sequencing verification and error localization. The complete encoding and decoding pipeline has been made publicly available on GitHub (https://github.com/Charlie‐99‐max/DNA‐Memory‐pic).

Encoded DNA data were cloned into information plasmids paired with specific fluorescent protein expression modules, thereby establishing a deterministic mapping between fluorescent signals and data content at the system level. This mapping remained invariant throughout the file lifecycle, enabling subsequent retrieval to be performed solely based on physical signals without reliance on PCR amplification or sequence recognition.

4.5. Preparation and Characterization of Engineered Living Memory Microspheroids (ELMMs)

ELMMs were fabricated using a droplet‐based microfluidic approach. The microfluidic system consisted of syringe pumps (DK Infusetek Co., Ltd., Shanghai, China), a flow‐focusing microfluidic chip, a low‐temperature crosslinking region, and a digital microscope (Shenzhen Supereyes Co., Ltd., Shenzhen, China) for droplet generation, real‐time monitoring, and crosslinking control.

Engineered Escherichia coli cells in the exponential growth phase were collected by centrifugation and resuspended in 7.5% (w/v) gelatin (Sigma–Aldrich, porcine skin) dissolved in PBS as the dispersed phase. Novec 7500 (Techu Scientific, Tianjin, China), containing 0.1% Pico‐Surf surfactant, was used as the continuous phase to stabilize droplet interfaces. Flow rates of both phases were precisely controlled using syringe pumps.

The dispersed phase was injected into the continuous phase through a flow‐focusing microfluidic chip with a 50 µm constriction, generating highly monodisperse water‐in‐oil droplets. The droplets subsequently entered a low‐temperature solidification channel, where gelatin underwent thermoreversible physical gelation under cooling, forming structurally stable ELMMs. After droplet collection, 100 µL of demulsifier (Emulseo, Bordeaux, France) was added, followed by mixing and centrifugation at 5000 rpm for 5 min. The lower oil phase was removed by pipetting.

The microspheroids were washed twice with 1× PBS (Jiangsu Femeng Biotechnology Co., Ltd.) to remove residual oil and filtered through a 70 µm membrane (Shandong Barlow Biotechnology Co., Ltd.). Processed ELMMs were resuspended in bacterial cryopreservation solution (Coolaber, SL9291), subjected to gradual cooling, and stored at −80°C or directly used for subsequent experiments.

ELMM size distribution, circularity, and fluorescence signal distributions were systematically characterized by microscopy and flow cytometry. Thermally triggered release behavior was evaluated by transferring ELMMs into 2×YT medium containing ampicillin (100 µg·mL^{− 1}; Sangon Biotech, Shanghai, China) and increasing the temperature to 37°C. Bacterial growth curves after release were used to assess the impact of encapsulation–release processes on cell viability and information stability.

4.6. Construction, Lyophilization, Storage, and Rehydration of Living Disks

Living Disks were constructed by physically mixing ELMMs carrying different information fragments and loading them as intact ensembles. Specifically, mixed ELMMs were either loaded into storage chambers of microfluidic chips (Namocell) or placed directly into 1.5 mL centrifuge tubes (Sigma‐Aldrich), forming file‐level storage assemblies.

Prior to lyophilization, ELMM mixtures were resuspended in a protective solution containing 0.4 M trehalose (Sigma‐Aldrich) and incubated at 4°C for at least 30 min to ensure sufficient penetration of the cryoprotectant into the ELMM. Samples were then rapidly frozen in liquid nitrogen for 10 min and immediately transferred to a FreeZone 6L benchtop freeze dryer (Labconco, Kansas City, USA) for vacuum lyophilization for no less than 20 h. During lyophilization, the condenser temperature was maintained at −80°C, and chamber pressure was controlled below 0.2 mbar.

After lyophilization, chamber pressure was gradually equilibrated by slow gas backfilling. Samples were rapidly transferred to sealed bags containing desiccants and handled in a nitrogen glove box under low‐humidity and low‐oxygen conditions, followed by vacuum sealing in aluminum foil pouches. This workflow enabled whole‐database lyophilization of Living Disks as intact entities rather than discrete processing of individual molecules or cells.

Lyophilized Living Disk samples were stored in glass desiccators (SGRG, diameter 150 mm) at room temperature. Storage temperature was maintained at 25°C ± 5°C, and relative humidity was controlled at 20% – 30%. For rehydration, lyophilized samples were first exposed to a humid environment at 4°C, followed by gradual addition of 2 × YT medium with gentle agitation, completing rehydration within approximately 10 min. Rehydrated ELMMs were re‐evaluated for structural integrity, fluorescence signal intensity, and thermal response behavior to verify the stability and usability of Living Disks under the tested ambient dry‐storage and repeated freeze‐drying–rehydration conditions.

4.7. Construction of the Living Drive and Automated Regeneration and Replenishment Workflows

The Living Drive module was designed to enable automated regeneration and database replenishment of retrieved files, while providing a modular interface for future integration of information‐updating operations. In the current prototype, the automated Living Drive workflow includes thermo‐responsive release, bacterial culture, aqueous‐ink preparation, droplet generation, and ELMM re‐encapsulation. CRISPR‐mediated information replacement was experimentally validated as an offline single‐round information‐plasmid cassette‐replacement process and was not integrated into the automated Living Drive workflow. All 3D design files (STL) and material lists have been uploaded to GitHub and made openly accessible (see Supporting Information) to ensure reproducibility.

The Living Drive chip was fabricated using high‐precision SLA 3D printing (Form 3+, Formlabs Inc.) with biocompatible BioMed Clear Resin V5. The chip integrated a spindle‐shaped static culture chamber with a volume of approximately 300 µL, along with printed bases and curing channels. To eliminate potential cytotoxicity from residual monomers, printed chips were sequentially subjected to isopropanol ultrasonic cleaning, UV post‐curing, and dynamic soaking in sterile PBS on a shaker for 72 h, with PBS replaced every 12 h.

To suppress biofilm formation and nonspecific bacterial adhesion during long‐term static culture, microchannels were pre‐treated with 1% (w/v) Pluronic F‐127 (Sigma–Aldrich) for 1 h, followed by thorough rinsing with deionized water. Microfluidic chips used in experiments were purchased from Anhui Chixin Biotechnology Co., Ltd. (mixing chip: LNP C02; encapsulation chip: DT‐50).

During regeneration, retrieved target ELMMs were directly delivered into the Living Drive module. After system assembly, the heating module was activated to maintain the culture chamber at 37°C for static cultivation. Under these conditions, the gelatin matrix of ELMMs melted, releasing encapsulated engineered bacteria as initial seeds for data replication. Bacteria proliferated within the integrated culture chamber, with cultivation time controlled according to a pre‐fitted growth model derived from empirical growth curves obtained in preliminary experiments, ensuring that cell concentrations exceeded the threshold required for re‐encapsulation (>2 × 10⁸ CFU·mL^{− 1}).

Upon completion of cultivation, bacterial suspensions were pneumatically driven into downstream mixing and droplet‐generation regions. In the mixing zone, bacterial suspensions and 15% gelatin solution were combined at a flow rate ratio of 500 µL·h^{− 1}: 500 µL·h^{− 1} to yield a final gelatin concentration of 7.5%. The mixture entered the droplet‐generation region, where it was sheared into monodisperse water‐in‐oil droplets using HFE‐7500 fluorinated oil containing 2% (w/w) fluorosurfactant (Emulseo, France) as the continuous phase at a flow rate of 4000 µL·h^{− 1}. Droplets were solidified in a low‐temperature crosslinking zone and collected in an ice‐water bath.

Following the collection, 100 µL demulsifier (Emulseo) was added, mixed, and centrifuged at 5000 rpm for 5 min to remove the oil phase. ELMMs were washed twice with 1×PBS (Jiangsu Femeng Biotechnology Co., Ltd.), filtered through a 70 µm membrane (Shandong Barlow Biotechnology Co., Ltd.), and finally resuspended in bacterial cryopreservation solution (Coolaber, SL9291). After gradual cooling, samples were stored at −80°C or used directly for subsequent experiments.

Regenerated ELMMs were systematically characterized for size distribution, circularity, and fluorescence signal distribution to evaluate reproducibility as standardized storage units. A subset of regenerated ELMMs could be replenished back into the Living Disk, while other regenerated cell populations could be used for downstream readout or offline single‐round CRISPR‐mediated information replacement. Thus, the current Living Drive established a closed‐loop operational paradigm of retrieval–regeneration–replenishment, with information updating validated as a compatible but not yet automated module.

4.8. File‐Level Physical Random Access Method

File retrieval in the Living Disk is based on the fluorescent tags carried by ELMMs, implemented via a fluorescence sorter. For integrated verification, the Namocell Pala (Bio‐tech) dual‐laser flow sorter was employed, while the Bigfoot spectral cell sorter was used for non‐integrated verification experiments. After rehydration, the Living Disk was loaded into a sorting chip, where the system read the fluorescence signals of ELMMs sequentially in single‐particle detection mode. Retrieval criteria were defined by preset multi‐channel fluorescence gating logic to distinguish target files from non‐target files. Except for the Namocell‐based integrated verification experiments, all flow cytometry analysis experiments were performed using the Bigfoot spectral cell sorter (Thermo Fisher Scientific) equipped with a standard 150 µm nozzle. Initial gating was established based on fluorescein isothiocyanate (FITC) channel signals, as well as forward scatter (FSC) and side scatter (SSC) signals. Fluorescence signals for enhanced green fluorescent protein (eGFP), monomeric red fluorescent protein (mRFP), and tag blue fluorescent protein (TagBFP) were acquired using 488, 561, and 405 nm excitation lasers, respectively. Since microspheroids encapsulating different Escherichia coli strains generated specific fluorescent signals, further gating and screening were performed based on fluorescence type and intensity. All flow cytometry experiments were conducted at 4°C with a fixed detection volume of 600 µL (for both Namocell and Bigfoot), and each experimental group included three biological replicates. Data analysis was performed using FlowJo v.10.8.1 software. Each sorting event was designed to isolate a single ELMM as the minimal physical retrieval unit. Single‐droplet sorting assays were used to quantify the practical occupancy and precision of this operation, rather than to assume perfect single‐ELMM isolation in every event. The purity and information accuracy of the sorted ELMMs were verified through microscopic imaging, secondary flow cytometry detection, and subsequent culture and sequencing analysis after bacterial release.

4.9. Definition and Measurement of System Performance Metrics

To quantitatively evaluate the file‐level operational performance of the Living Disk–Drive system, key metrics including sorting‐stage retrieval rate, architecture‐level storage‐density estimates, and non‐destructive access capability were defined and interpreted under explicitly stated boundary conditions. Unless otherwise specified, performance analyses used a single ELMM as the minimal physical retrieval and statistical unit.

The sorting‐stage retrieval rate was defined as the amount of DNA‐encoded file payload associated with ELMMs physically sorted per unit time. For the current proof‐of‐concept system, this value was calculated using the experimentally encoded file size, the number of ELMMs required per file, and the experimentally relevant sorting‐output rate of the Namocell system. Manufacturer‐specified enrichment or sample‐processing event rates were treated only as upper‐bound instrument‐level estimates and were not interpreted as experimentally achieved end‐to‐end retrieval bandwidth.

Storage‐density metrics were estimated under explicitly stated model assumptions rather than measured as optimized device‐level density. We distinguished the corrected dried‐particle architecture‐level density from the apparent operational density of the current prototype, which includes rehydrated sample volume, chip‐handling volume, reservoir and tubing dead volume, particle packing, aggregation, and retrieval‐accessibility penalties. These estimates were used to define a quantitative framework for future device optimization, rather than to claim that the current chip‐interfaced prototype had achieved the theoretical density ceiling.

Non‐destructive access capability was evaluated using a combination of fluorescence‐based physical retrieval, post‐retrieval bacterial regrowth, re‐encapsulation, repeated lyophilization–rehydration assays, and repeated Living Drive release–regenerate–re‐encapsulation cycles. These assays validate key modules of repeated operation under the tested conditions, but they do not yet constitute a long‐term, fully integrated multi‐cycle database‐composition stability test or a multi‐cycle rewrite‐endurance benchmark.

4.10. Bacterial Culturing, Plasmid Extraction, Sequencing, Imaging, and Statistical Analysis

4.10.1. Plating and Culturing of Bacteria and ELMMs

Under sterile conditions, bacteria released from sorted ELMMs or bacterial suspensions obtained from other experiments were spread onto LB agar plates supplemented with appropriate antibiotics. After plating, the agar plates were incubated at 37°C for 36 h. Colony‐forming units (CFUs) were counted to quantify bacterial viability.

The formation and growth of colonies on agar plates were used to directly assess the survival of encapsulated bacteria after sorting, regeneration, and related operations. These measurements provided quantitative evidence for the stability of the ELMM encapsulation process and the reliability of the file‐level sorting workflow.

4.10.2. Extraction of Bacteria from ELMMs

Sorted ELMM samples were diluted to an appropriate concentration and directly plated onto solid culture media, followed by incubation at 37°C for approximately 36 h. After incubation, well‐isolated bacterial colonies were observed on the agar surface. Individual colonies were picked and inoculated into LB liquid medium containing the corresponding antibiotics, followed by overnight cultivation at 37°C under shaking conditions.

After cultivation, bacterial suspensions were transferred to centrifuge tubes and pelleted by centrifugation. The resulting bacterial pellets were used for subsequent plasmid extraction and sequencing analyses.

4.10.3. Extraction of Plasmid DNA from Bacteria

Plasmid DNA was extracted from Escherichia coli DH5α using a standard alkaline lysis method. Briefly, bacterial pellets were resuspended in Solution I (containing glucose, Tris‐HCl, and EDTA), followed by the addition of Solution II (containing NaOH and SDS) to lyse the cells and release nucleic acids. After gentle mixing, Solution III (containing potassium acetate and acetic acid) was added, and the mixture was inverted gently to precipitate proteins and genomic DNA.

Following centrifugation, the supernatant was transferred to a new tube and mixed with an equal volume of phenol/chloroform. After vigorous mixing and centrifugation, the aqueous phase was collected, and plasmid DNA was precipitated by adding two volumes of cold absolute ethanol and incubating at −20°C. The DNA pellet was washed with 70% ethanol, air‐dried, and resuspended in TE buffer or nuclease‐free water. All procedures were performed at low temperature to minimize DNA degradation. Plasmid extraction and sequencing services were performed by Sangon Biotech (Shanghai, China).

4.10.4. Sequencing Data Quality Control and Sequence Analysis

Raw sequencing reads were processed using a three‐step quality control pipeline. First, adaptor sequences were removed using Cutadapt (v1.2.1). Second, low‐quality bases (Q < 20) were trimmed from the 3′–5′ ends of reads using PRINSEQ‐lite (v0.20.3). Third, chimeric sequences were removed using USEARCH (v11.0.667) in de novo mode with default parameters. The remaining high‐quality reads were used for downstream analyses.

Target sequences were extracted using a custom Python script, which retained reads only when both forward and reverse primer sequences were perfectly matched. Unique sequences and total read counts were then calculated. The top 20 most abundant sequences were selected for multiple sequence alignment against reference sequences for mutation analysis. Data visualization and statistical plotting were performed using custom R scripts.

4.10.5. Sequencing of Retrieved Files

Sanger sequencing: Individual bacterial colonies retrieved after file access were subjected to Sanger sequencing to determine the DNA sequences carried by each clone (performed by Sangon Biotech, Shanghai, China). Sequencing results were aligned to the original reference library to calculate retrieval accuracy. Multiple independent colonies were sequenced to ensure reproducibility and reliability. This analysis was used to verify the fidelity and stability of DNA data during encoding, storage, and retrieval.

Next‐generation sequencing (NGS): High‐throughput sequencing was performed on bacterial populations subjected to long‐term serial passaging (e.g., up to 216 generations) or after information rewriting. Following PCR amplification, sequencing libraries specifically targeting the information‐encoding regions were prepared. Consensus sequence analysis was conducted to quantify the weighted average error rate and the fraction of error‐free reads, thereby assessing the stability of stored information during prolonged passaging and after rewriting.

For library preparation, PCR products were sent to Sangon Biotech (Shanghai, China). A two‐step PCR protocol was used. In the first PCR, reactions (25 µL total volume) contained 2 µL DNA (10 ng µL^{− 1}), 1 µL forward primer mix (10 µM), 1 µL reverse primer mix (10 µM), and 15 µL 2× KAPA HiFi ReadyMix. Amplification was performed on a BIO‐RAD T100 thermal cycler using the following program: 98°C for 5 min, followed by 8 cycles of 98°C for 30 s, 60°C for 30 s, and 72°C for 30 s, with a final extension at 72°C for 5 min, then held at 4°C.

PCR products were verified by electrophoresis on 1% (w/v) agarose gels in TBE buffer stained with SYBR Green I and visualized under UV illumination. Amplicons were purified using AMPure XP beads before the second PCR. The second PCR (30 µL total volume) contained 2 µL DNA (10 ng µL^{− 1}), 1 µL indexed universal P7 primer (10 µM), 1 µL indexed P5 primer (10 µM), and 15 µL 2× KAPA HiFi ReadyMix. Cycling conditions were 98°C for 3 min, followed by 5 cycles of 94°C for 30 s, 55°C for 20 s, and 72°C for 30 s, with a final extension at 72°C for 5 min.

After purification with AMPure XP beads, libraries were quantified, pooled, and subjected to paired‐end sequencing (PE150 or PE300) on Illumina NovaSeq 6000 or MiSeq platforms (Illumina, San Diego, CA, USA).

4.10.6. Scanning Electron Microscopy (SEM) Imaging

ELMM samples were fixed overnight at 4°C in a fixation solution containing 2% paraformaldehyde and 2.5% glutaraldehyde. Samples were washed three times with 0.1 M PBS at 4°C, followed by graded ethanol dehydration using 30%, 50%, 70%, 80%, 90%, and 95% ethanol, and then three washes with 100% ethanol.

Samples were subsequently transferred to a 1:1 (v/v) ethanol/tert‐butanol solution once, followed by two washes in pure tert‐butanol. After rapid freezing in liquid nitrogen, samples were freeze‐dried using a Labconco FreeZone 6L freeze dryer. Dried samples were mounted on aluminum stubs using conductive adhesive, sputter‐coated with gold, and imaged using a scanning electron microscope.

4.10.7. Bright‐Field Imaging

Bright‐field imaging was performed using a Nikon C2 Plus confocal microscope equipped with 10× and 20× objectives. ELMM samples were placed on glass slides and covered with coverslips. Images were acquired under transmitted light illumination to visualize overall microbead morphology. Image processing and quantitative analysis of ELMM size and shape distributions were performed using NIS‐Elements and ImageJ software.

4.10.8. Confocal Fluorescence Imaging

Confocal imaging was conducted using an Olympus FV3000 laser scanning confocal microscope (Evident/Olympus, Tokyo, Japan). Rehydrated ELMM samples were excited using 488, 561, and 405 nm diode lasers to detect green (EGFP/sfGFP), red (mCherry/mScarlet3), and blue (TagBFP) fluorescence signals, respectively. Z‐stack images were acquired to reconstruct the 3D structure of ELMMs. Image acquisition and 3D reconstruction were performed using the FV31S‐SW software (Olympus), and further spatial distribution analysis of fluorescent bacteria within the microbeads was conducted using ImageJ.

4.10.9. Statistical Analysis

All statistical analyses were performed using GraphPad Prism 9.5.1 (La Jolla, CA, USA). Comparisons between two groups were conducted using two‐tailed unpaired Student's t‐tests, while comparisons among multiple groups were performed using one‐way analysis of variance (ANOVA) followed by Tukey's post hoc test. Data are presented as mean ± standard error of the mean (SEM). Statistical significance was defined as ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001; n.s., p > 0.05, not significant.

Author Contributions

Zhuo Xiong and Hao Luo conceived the study, with Zhuo Xiong supervising the overall research. Hao Luo designed and developed the droplet microfluidics platform based on a microinjection strategy; prepared, cryopreserved, analyzed, and sorted ELMMs; led the design and fabrication of the Living Drive; built the prototype Living Disk–Drive system; and performed the primary data processing and mapping analyses. JinKai Gao and TianYu Huang contributed to the design of dual‐plasmid rewriting engineered cells and participated in selected experiments and data analysis. ChengHao Cao and ZeYang Yu assisted in the design and fabrication of the Living Drive and prepared schematic illustrations. YingKai Xia assisted with Living Drive fabrication, and XiangXiang Huang assisted with the preparation, cryopreservation, analysis, and sorting of ELMMs. Hao Luo drafted the manuscript, and Zhuo Xiong and YongCong Fang revised it.

Funding Statement

This work was supported by the new faculty start‐up funding provided by Tsinghua University (grant number: 53330200321).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File 1: adma73806‐sup‐0001‐SuppMat.pdf.

Supporting File 2: adma73806‐sup‐0002‐MovieS1.mp4.

Supporting File 3: adma73806‐sup‐0003‐MovieS2.mp4.

Supporting File 4: adma73806‐sup‐0004‐MovieS3.mp4.

Data Availability Statement

Research data are not shared.