Suppression of errors in collectively coded information

TL;DR

This paper investigates how collective encoding in a virtual circular genome architecture can suppress errors and mutants, enabling robust information transmission even at high mutation rates, with implications for prebiotic evolution.

Contribution

It introduces a model of collective genome encoding that reduces mutation propagation, supported by experiments, simulations, and theoretical analysis.

Findings

Mutant oligomers are suppressed in collective encoding.

Competition for binding partners explains mutant suppression.

Robust information propagation is possible at high mutation rates.

Abstract

Modern life largely transmits genetic information from mother to daughter through the duplication of single physically intact molecules that encode information. However, copying an extended molecule requires complex copying machinery and high fidelity that scales with the genome size to avoid the error catastrophe. Here, we explore these fidelity requirements in an alternative architecture, the virtual circular genome, in which no one physical molecule encodes the full genetic information. Instead, information is encoded and transmitted in a collective of overlapping and interacting segments. Using a model experimental system of a complex mixture of DNA oligomers that can partly anneal and extend off each other, we find that mutant oligomers are suppressed relative to a model without collective encoding. Through simulations and theory, we show that this suppression of mutants can be explained by competition for productive binding partners. As a consequence, information can be propagated robustly in a virtual circular genome even at mutation rates expected under prebiotic conditions.

Equations

$$F_i^A\equiv \frac{{\sum }_{j\in D_i^A}{c}_{ss,j}}{{\sum }_{j\in B_i}{c}_{ss,j}},$$

$$f^A=\frac{1}{c_{tot}^A}\sum _{i\in O_A}{c}_{ss,i}{F}_i^A,$$

$$F_i^A\sim x_i^A/l_i,$$

Full text

Faithful copying of heritable information is a basic requirement for genomes. Standard accounts of fidelity emphasize enzyme-based mechanisms ranging from nucleotide selectivity[1] and exonucleolytic proofreading[2] to post-replicative mismatch repair[3] acting on a single, continuous template. These mechanisms are powerful but in extant biology are based on sophisticated protein machinery, and they must operate below the well-known error-catastrophe threshold to maintain information[4]. A complementary approach, often overlooked, involves changing the architecture of the genome, i.e., how genetic information is physically laid out. Biology offers many precedents for diverse architectures: the fragmented chromosomes and plasmids of Borrelia burgdorferi[5]; the linear, circular and branched forms of mitochondrial genomes[6]; the partitioning into a micronucleus and macronucleus in ciliates[7]; and the interlocked kinetoplast DNA networks of many parasites[8]. Yet the impact of such architectural choices on how errors arise and propagate remains poorly understood.

Alternative architectures for genomic information are also motivated by the search for minimal self-replicating systems, whether envisioned for synthetic biology or as models of early life. Copying an extended template end-to-end typically requires elaborate machinery that can unwind and manage long duplex regions; in modern cells this role is played by the multi-enzyme replisome (helicases, polymerases, gyrases and many other accessory factors). Consequently, even engineered minimal cells currently rely on sizable gene sets to support these processes[9]. Such machinery is unavailable in early evolutionary settings and in more minimal synthetic constructs. Hence it is natural to consider distributed architectures in which information is carried by many short genomic fragments and replication proceeds via many short, parallel extensions rather than a single long uninterrupted pass[10, 11, 12, 13, 14, 15, 16, 17, 18]. We refer to this class of architectures where a population of short, overlapping strands jointly stores and transmits a sequence as a collective encoding.

These considerations motivate the central question we address here: Do collectively encoded genomes, by virtue of their architecture alone, improve replication fidelity? We address this question in the context of the virtual circular genome (VCG) framework[19, 20], in which overlapping genomic fragments map to a circular consensus sequence, even though no single strand contains the whole. We use an experimental DNA-based implementation of a VCG, together with simulations and theory, to compare how wildtype and mutant sequences propagate in a collectively encoded pool.

Our results reveal an inherent asymmetry created by the architecture. Wildtype sequences that match the consensus are distributed across many different strands and hence retain many routes to continued copying. In contrast, mutations appear at singular locations and hence can have reduced routes to copying. The deficit is most severe for changes near the growing 3′ end, where the location both cuts off continuation options and increases the likelihood of stalling—premature halts that maroon mutated pieces in short, non-contributing products—consistent with prior observations and models of stall-mediated error suppression and its possible role as a precursor to proofreading[21, 22, 23, 24]. Consequently, mutants are intrinsically suppressed relative to wildtype, with the strongest effects for 3′-terminal changes.

Our work shows that by distributing information across interacting oligomers, collective encoding provides an architecture-level route to fidelity that complements enzymatic error correction. This alternative is especially pertinent to early life forms, where elaborate enzymes may be absent, and to synthetic self-replication, where robust information transmission from simple parts is a design goal[25, 26, 27].

I Experimental model

We designed a DNA-based system composed of short, overlapping oligomers (oligos) that form a redundant circular architecture. The full VCG comprises 12 distinct 25 nt oligos and their reverse complements. Each of the 12 oligos in one direction overlaps with its downstream neighbor (i.e., the next oligo in the 5′–3′ direction) by 20 bases, encoding a 60 bp circular genome (Fig. 2A, Supp. Table I). Thus, each oligo initially has four binding partners that productively allow it to be extended or part of it to be copied, i.e., reverse-complementary oligos that can hybridize and serve as templates and/or as primers, with partial complementarity ranging from 5 to 20 bp. Each oligo also has an additional five binding partners which anneal in unproductive configurations that do not allow for primer extension. Less structured DNA experimental models with pools of oligos have been investigated in prior work[14, 16, 28, 29, 12, 15, 13].

To evaluate how mutations propagate within this architecture, we introduced a 25 nt mutant oligo that differs from one VCG oligo by a single 4 bp substitution near the center of the sequence (Fig. 2B, Supp. Sec. I). Mutant oligos were spiked into VCG mixtures at defined low proportions (2.5%, 5%, or 25% relative to the wildtype) to mimic rare variant emergence.

Extension of VCG oligos was driven by thermal cycling with Bst DNA Polymerase (Large Fragment, NEB M0275), using 10–30 cycles of alternating denaturation ($80~^{\circ}\mathrm{C}$) and annealing/extension ($35~^{\circ}\mathrm{C}$), followed by a final enzyme deactivation step ($90~^{\circ}\mathrm{C}$) for 10 minutes. Notably, no new oligomers are supplied to this system; extension is initiated by overlap-driven annealing of partially complementary VCG oligos (Supp. Sec. III). Gel electrophoresis confirmed that initial 25 nt VCG oligos were extended incrementally over cycles, with products reaching a length of $\sim$45 nt by cycle 10, and further reaching $\sim$60 nt by cycle 20, with gel profiles remaining unchanged through cycle 30 (Fig. 2D). Duplexes up to 45 bp ($T_{\mathrm{m}}=80~^{\circ}\mathrm{C}$) likely remain partially meltable during cycling, allowing oligo reshuffling and continued extension, whereas  60 bp duplexes ($T_{\mathrm{m}}=82~^{\circ}\mathrm{C}$) are too stable to denature, halting further growth. We refer to this final state as ‘pool stasis’.

To quantify the relative amplification of wildtype and mutant alleles, we used sequence-specific qPCR with forward primers that selectively bind either the wildtype or mutant sequence, combined with a shared reverse primer (Fig. 2C, Supp. Sec. IV). Two aliquots from each thermocycled sample were analyzed in parallel using both primer sets, allowing independent quantification of wildtype and mutant allele concentrations (Fig. 2D–E). Ct values were determined from the qPCR amplification curves and converted into absolute concentrations using standard curves generated from known DNA concentrations included in the same qPCR run. Control experiments confirmed that the qPCR assay reliably quantifies each allele sequence with high specificity and within the relevant concentration range (Supp. Fig. S1).

qPCR analysis revealed that the wildtype allele was consistently amplified more than the mutant allele over 30 cycles of VCG extension; competitive oligo interactions favor wildtype proliferation and effectively suppress the mutant (Fig. 2E, Supp. Sec. VI). Notably, this suppression effect is time-dependent: before cycle 10 the mutant is strongly suppressed, while from cycles 10 to 20, wildtype and mutant grow with equal speed until pool stasis. We define a ‘wt advantage’ metric as the ratio of wildtype and mutant amplification at pool stasis (i.e., at cycle 30).

To investigate the role of VCG architecture in mutation suppression, we varied the system’s ‘virtualness’ $V$ – defined as the number of overlapping oligo pairs encoding the genome – by constructing VCGs with 12, 6, or 3 oligo pairs (Fig. 3A, Supp. Table II). All of these sets of oligos encode the same 60 bp genome but differ in the number of distinct potential reverse complement binding partners available to each oligo. Lower virtualness (e.g., the 3-oligo VCG) more closely resembles a real physical circular genome in that it has low redundancy measured by how many different oligos cover a given subsequence of the genome. We observed stronger suppression of mutant amplification in higher-virtualness VCGs across all mutant input levels we tested, including 2.5% (Supp. Fig. S4), 5% (Fig. 3A; Supp. Fig. S5), and 25% (Supp. Fig. S6). In the 12-oligo VCG, the wildtype outcompeted the mutant by over 10-fold, while in the 3-oligo VCG, the amplification advantage for the wildtype was nearly absent (Fig. 3B).

We next examined how the position of the 4 bp mutation within the mutant oligo affects its suppression. Using additional mutant variants with mutations placed near the 3′ or 5′ ends, we found that mutations near the 3′ end were most strongly suppressed during extension (Fig. 3C–D). In contrast, the 5′ end mutant exhibited amplification levels nearly indistinguishable from the wildtype. This trend held consistently across all VCG virtualness levels and initial mutant proportions (Fig. 3D, Supp. Fig. S4–6), suggesting that suppression is more effective when the mutation is closer to the 3′ end—the site where oligo extension initiates. This positional bias aligns with the expectation that errors or mismatches near the 3′ end would exhibit a decreased likelihood of propagation.

Together, these findings demonstrate that both VCG virtualness and mutation position govern selective replication within this oligo-extension-based system. The results suggest that VCG architectures can impose intrinsic fidelity constraints, naturally suppressing the propagation of mutant alleles relative to the wildtype allele during thermal cycles.

II Simulation

II.1 Simplified VCG Model

To understand the source of position- and virtualness-dependent mutation fates in the VCG, we build a simplified model of VCG replication (Fig. 4A). In this model, we consider an initial pool of oligos which grows through repeated thermal cycles. This initial oligo pool is based on the experimental oligo pools studied in the previous section. It contains 12 oligo sequences of length 25 nt which form a consensus VCG sequence of length 60 bp with an offset of 5 bp between each consecutive oligo. In addition, the pool also contains all the reverse complements of the original 12 oligos, plus a small concentration of an additional oligo that is identical to the underlying VCG sequence except at a single position, where it contains a mutant allele. This is in contrast to the experimental mutant oligo, which differs from its wt counterpart in a 4 nt block.

The initial pool undergoes thermal cycles which consist of the following:

First, oligos anneal with any possible complementary partner to form duplexes, following irreversible second-order reaction kinetics until no further annealing reactions are possible. Only oligos which have exact contiguous overlaps above a threshold $o_{min}$ amount are allowed to bind. Here $o_{min}=2$. 2. 2.

Second, if an oligo has an annealed 3′ end, it acts as a primer and is extended using its annealed duplex partner as a template. This extension continues without any errors until the extending oligo reaches either the 5′ end of its template or a maximal length $l_{max}$ equal to the length of the consensus VCG sequence (60 nt). 3. 3.

Finally, during the melting stage, duplexes dissociate into single-stranded oligomers unless their overlap exceeds a specified threshold length $o_{max}$. Here $o_{max}=55$, as approximately observed in experiments.

We note that annealing in any single thermal cycle is random in that the annealed duplexes are kinetically determined. Thus, only a small fraction of an oligo may be bound to its perfect complementary pair. For instance, in the initial oligo pool where oligos are of length 25 nt and are separated by offsets of 5 nt, then any oligo has $9$ potential binding partners, only one of which is its perfect complementary pair. Binding the perfect complementary pair forms a blunt-end duplex which does not allow for further extension. However, each oligo has an additional 8 other binding partners which can bind in equal probability; 4 of those 8 bind in duplexes which allow for extension.

Therefore, because this system anneals through irreversible kinetics, not all pairs bind with perfect complementarity and hence extension can occur through the formation of duplexes with overhangs. The formation of duplexes with perfect complementarity is further disrupted by the melting phase of thermal cycles on a timescale $\tau_{cycle}$. For more details on how our model is defined and implemented, please see Supp. Sec. VII.

II.2 Mutant suppression in the simplified model

Following simulation of repeated thermal melt-anneal-extend cycles, we first observe that oligos extend on each other (Fig. 4B, top); the distribution of oligo lengths shifts from their initial values (25 nt) to a maximal length set by a combination of $o_{max}$ (the maximal duplex overlap length for unbinding during the melt phase) and $l_{max}$ (the maximal oligo length). As a consequence of their extension, oligos become longer and more likely to be trapped in duplexes that overlap too much to unbind during the melting step. In particular, while oligos at the end of thermal cycle 0 are still all single-stranded, virtually all oligos are bound in duplexes of length 55–60 bp at the end of cycle 10 (Fig. 4B, bottom). We refer to this state where all oligos are duplex-bound as ‘pool stasis’, since the VCG cannot replicate anymore. In simulations, we see pool stasis occurs at approximately cycle 10.

Extension also allows for an increase in the concentration of mut and wt alleles, despite the fact that no new oligos are created in our model. As oligos extend off each other as templates, new strands bind downstream of the mut and wt alleles (i.e., between the allele on an oligo and the 3′ end of that oligo) and extend past those alleles, thereby increasing the concentration of oligos which contain the two alleles.

To model the qPCR readouts of the DNA model experiments, we track the relative amplification of the mut and wt alleles by measuring the total concentration of oligos that contain the allele along with a flanking region around the allele. This flanking region is set by the oligo that the mut allele initially appears on. After normalizing these concentrations to the initial concentration of flanking region oligos, we can define a ‘wt advantage’ statistic as the ratio between the wt region amplification at pool stasis and the mut region amplification at pool stasis. We find that the wt advantage increases with the virtualness $V$ of the initial oligo pool (Fig. 4C), where $V$ here can be defined as the number of unique oligo pairs that cover the consensus VCG sequence in the oligo pool (either 3, 6, or 12). We also find that the wt advantage measured at pool stasis is stronger as the initial mutant position shifts from the 5′ end to the 3′ of the oligo it starts on (Fig. 4D).

II.3 Collective binding partner effects predict mutant suppression

In order to find an explanation for the observation of mutant suppression both in our experimental DNA system as well as our simplified simulation model, we focus on the case where the mutant allele is initially at the 3′ end of the oligo on which it occurs. This is both the situation where the mutant suppression is the most dramatic (Fig. 4D), as well as the most relevant for mutant alleles that arise through mistakes during extension.

We propose that the wt advantage comes from a binding partner effect. In the VCG pool, each oligomer has multiple potential reverse complement binding partners, with the number of potential partners increasing with the virtualness $V=3,\ldots,12$ of the initial VCG oligo pool. These binding partners all compete for annealing with one another on the template oligo, on equal footing with the perfect reverse complement partner.

Crucially, only binding partners whose 3′ ends anneal downstream of the allele (i.e., between the allele and the 3′ end of the template), can increase that allele’s reverse complement concentration via extension. In contrast, binding partners whose 3′ ends are upstream of the allele may extend but do not result in copying of the allele into its reverse complement, precluding an increase in concentration of the allele itself upon further rounds of extension.

Therefore, a mut allele initially located near the 3′ end of an oligo has fewer binding partners that can bind downstream of that allele and make its reverse complement during extension. In contrast, wt alleles are initially distributed across many different oligo locations. When the wt allele is located towards the 5′ end of a template oligo, it can be copied by almost all binding partners that anneal to that template (see Fig. 5A).

The binding partner effect implies that the instantaneous per-cycle rate at which an allele on a given oligo can produce more copies of its reverse complement is set by two quantities: 1. the concentration of ‘productive’ binding partner oligos which can bind and extend past the allele and are thus productive in copying that allele, and; 2. the concentration of all binding partner oligos that compete with (1), regardless of whether they can pick up the allele upon extension. An allele is more productive if it exists further from the 3′ oligo end and hence more of the oligo’s binding partners can form pairings that allow the allele to be copied. Correspondingly, an allele is also more productive if it sits closer to the 5′ oligo end since there are fewer binding partners that do not pick up the allele upon extension.

To formalize this intuition, we can define a productivity factor $F^{A}_{i}$ for an allele $A$ on a single-stranded oligo $i$ as:

[TABLE]

where $c_{ss,j}$ is the concentration of single-stranded oligo $j$, $D^{A}_{i}$ is the set of oligos that bind with 3′ ends downstream of the allele $A$ on oligo $i$, and $B_{i}$ is the set of all oligos that can bind to oligo $i$ regardless of the configuration of the duplex they form. (When oligo $i$ only exists in duplexes, $F^{A}_{i}$ is defined to be 0.)

Given this per-oligo definition of allele productivity, we can go further and integrate productivity across all oligos that carry that allele. Specifically, to compute the rate at which an allele produces its reverse complement per-cycle and per-oligo, we take a concentration-weighted average of the individual oligo productivity factors $F^{A}_{i}$ and define a productive pairing fraction $f^{A}$:

[TABLE]

where $O_{A}$ is the set of oligos that contain the allele $A$ and $c^{A}_{tot}$ is the total concentration of oligos containing $A$ regardless if they are single-stranded or in a duplex.

This definition captures the intuition of how the position of an allele on an oligo influences the allele’s productivity. For instance, when oligos are shorter than the oligo length maximum $l_{max}$, an allele $A$ coded on the 5′ end of oligo $i$ will have a maximal productivity factor $F^{A}_{i}=1$. This high productivity comes because every binding partner of oligo $i$ is capable of copying $A$ through extension, so every binding partner oligo index $j\in D^{A}_{i}$ is also $j\in B_{i}$. In contrast, $F^{A}_{i}=0$ for that same allele $A$ if it were coded at the 3′ end of oligo $i$; in this case, there are no downstream binding partners for $A$ on $i$, so $D^{A}_{i}$ is empty. An approximation of the link between allele position and productivity that interpolates between the 3′ and 5′ end location limits is therefore:

[TABLE]

where $x^{A}_{i}$ is the distance of allele $A$ from the 3′ end of oligo $i$, and $l_{i}$ is the length of oligo $i$; more generally, $F^{A}_{i}$ is a monotonically increasing function of $x^{A}_{i}$ which is 0 at $x^{A}_{i}=0$.

Therefore, in order to understand the time evolution of $f^{A}$ for the wt and mut alleles, we should track the distributions $P^{mut}(x)$ and $P^{wt}(x)$ of an allele’s distance $x$ from the 3′ end of oligos across single-stranded oligos in the oligo pool (Fig. 5B). While the mutant allele is initially 3′-proximal (concentrated at 5 nt from the 3′ end), by cycle 10 $P^{mut}(x)$ eventually spreads out. $P_{wt}(x)$, in contrast, is by construction initially more spread out since the wt allele was already encoded on a variety of different oligos (e.g. see Fig. 2A), and remains uniformly spread after all 10 cycles. Tracking the average distance $\langle x\rangle$ of the mut and wt alleles from the 3′ end (Fig. 5C, left) reveals that the mutant allele is initially closer to the 3′ end of oligos compared to the average wt allele, indicating an initial disadvantage in replicative potential. However, as the oligo pool replicates, the relative positions of the mean wt and mean mutant allele converge.

Indeed, if we calculate the time-dependent productive pairing fraction $f^{A}$ (Fig. 5C, right), we find that mut alleles are initially at a disadvantage, with a lower productive pairing fraction $f^{mut}$ compared to wt allele $f^{wt}$. This disadvantage decreases over the first few replication cycles, consistent with the decreasing gap between mean wt and mut distances from oligo 3′ ends. By cycle 5, the productive pairing fraction $f^{A}$ of wt and mut alleles are essentially equal and both very close to 0. The collapse of the productive pairing fraction $f^{A}$ for both wt and mut alleles is due to the fact that most alleles are bound in fully-extended blunt end duplexes. Such duplexes cannot participate in subsequent rounds of melt-anneal-extend and hence cannot contribute to the production of either wt or mut alleles.

In summary, the initial disadvantage faced by 3′-proximal mutant alleles due to the binding partner effect is transient. However, the cumulative impact of this early disadvantage leads to a persistent suppression of the mutant allele relative to the wildtype, even by the time of pool stasis. We next discuss how the stalling effect[22, 30, 31, 21, 32, 24, 33, 23, 34, 35] can significantly prolong this transient phase, thereby amplifying the overall suppression of the mutant allele.

III Stalling enhances collective suppression of mutants

The stalling effect describes the reduced rate of extension from a primer that ends in a mismatched base pair, compared to extension from a perfectly matched primer. Here, we investigate how stalling influences the propagation of mutant alleles within the VCG framework.

We consider annealed configurations where one of the oligos in the duplex is not correctly matched to its template at its 3′ end. This can either happen when an oligo with a mutant allele at its 3′ end is annealed to an oligo with a wt allele at that position, or it is an oligo with a wt allele at its 3′ end annealed to an oligo with a mutant allele there. For such duplexes where one of the oligos has a mismatch at its 3′ end, stalling delays extension by a time $\tau_{stall}$. Whether or not the duplex gets to extend therefore depends crucially on whether thermal cycling (with timescale $\tau_{cycle}$) drives the duplex apart before the stalling delay can be overcome. For primers with no 3′ mismatch, there is no barrier to extension and such primers will fully extend during the current thermal cycle. To model the effect of stalling, we posit that primers with a mismatched 3′ end get to extend through an exponential decay process with timescale $\tau_{stall}$. Hence by the end of a melt-anneal-extend cycle of length $\tau_{cycle}$, only a fraction $1-e^{-\frac{\tau_{cycle}}{\tau_{stall}}}$ of primers with a mismatched 3′ end will have extended (Fig. 6A). Please see Supp. Sec. VII for further discussion and implementation details on stalled extension.

Stalling can have a dramatic effect on the suppression of mutant alleles in a VCG scenario depending on the relative balance of $\tau_{stall}$ and $\tau_{cycle}$ (Fig. 6B, top). When $\tau_{stall}\ll\tau_{cycle}$, we find the expected mutant suppression described earlier, which is due to the transient replicative ability reduction arising from the binding partner effect. For intermediate $\tau_{stall}\sim\tau_{cycle}$, that transient period is greatly increased for the mutant because only a fraction of mismatched primers extends in any cycle; as a consequence, the mutant allele stays encoded at the 3′ end for many more cycles and has a reduction in replicative ability. When $\tau_{stall}\gg\tau_{cycle}$, the mutation remains confined to the 3′ end and there is no increase in mutant allele concentration at all. Thus stalling provides a suppression of mutant alleles by lengthening the transient period of reduced replicative ability for the mutant. We calculated the amplification of mutant alleles at the end of 10 thermal cycles as a function of the ratio $\frac{\tau_{stall}}{\tau_{cycle}}$ (Fig. 6B, bottom). At $\tau_{stall}=0$, stalling is not present, but mutant amplification is still suppressed due to collective VCG effects. Mutant amplification drops rapidly as a function of increasing $\tau_{stall}$.

IV Discussion

Our results show that collective replication in a virtual circular genome (VCG) architecture intrinsically suppresses the propagation of mutant alleles, even in the absence of high-fidelity copying or kinetic proofreading. Experiments, simulations, and a reduced theory show that suppression is strongest when mutations are 3′-proximal and increases with the VCG’s ‘virtualness’ (redundancy of overlapping oligos).

Intuitively, collective encoding of information builds a directional bias into replication. Alleles that match the consensus retain multiple routes to be replicated, whereas sequence changes (e.g., a mutant allele) prune those routes. We quantified these routes in terms of binding partners since an allele’s rate of being copied depends on how many annealed partners place their 3′ end downstream of that allele so that extension traverses it. 3′-proximal mutants have fewer such productive partners and thus are at a disadvantage compared to wildtype alleles that are coded at varying distances from the 3′ end. Stalling at mismatched 3′ ends prolongs this period of lowered replicative ability by delaying extension past the mismatch. Consequently, the combined effect of partner limitation and stalling leads to strong suppression that strengthens with virtualness and with increasing ratio $\tau_{stall}/\tau_{cycle}$ of stalling time to melt-anneal-extend cycle time. Unlike mispriming suppression in PCR, which relies on reduced binding affinity from mismatches, our effect persists even if mismatches have negligible impact on binding affinity.

We focused on the maintenance of information in a structured pool, although prior works often focus on the emergence of catalytic cycles from random sequence pools[13, 14]. Prior classic theories of distributed replication (e.g., hypercycles[4]) posit cooperative autocatalytic networks that can maintain information, but they have been difficult to realize at the molecular level[36]. In contrast, our work begins from a concrete molecular specification — a set of overlapping oligos and their allowed anneal–extend interactions — which makes the system experimentally tractable and tunable. This molecular grounding, rather than an abstract mutually catalyzing graph, allowed us to quantify architecture-driven fidelity. In this aspect, our system shares features with molecularly-defined experimental models[37, 38, 39, 40, 41, 42].

We expect the core ecological mechanism revealed by our DNA model to generalize beyond the specific conditions studied here, though the quantitative outcomes will depend on the replication chemistry (e.g., in the context of RNA[10, 20]). Differences in persistence length, strand displacement rates, non-enzymatic extension kinetics will shift quantitative behavior. While we focused on primer extension, templated ligation is natural in distributed architectures [13, 14, 15, 43, 44, 12, 45, 46, 47] and enables growth from both 5′ and 3′ ends. The binding partner effect will have to be studied and quantified in the context of ligation to understand biases intrinsic to that mode of replication.

In addition to copying errors during primer extension, several other processes can introduce mutations that affect VCG stability. Mismatched annealing[48] can generate transient mispairs that alter extension outcomes, while sequence repeats[49] may create novel annealing registers, allowing oligomers to extend from unintended regions and producing chimeric products. Chemical lesions such as deamination[50] represent another important source of mutation, and these are not confined to the 3′ terminus. Our systematic results on mutation position provide a framework to predict the fate of these diverse error sources within distributed architectures.

Our experiments emphasize primer extension in a fixed pool of oligomers that eventually become unproductive by forming long, unmeltable duplexes; sustained replication will require continual regeneration of short oligos (e.g., de novo synthesis from activated nucleotides[19], or cleaving oligos into short oligos[51]). Moving to driven continuous growth may strengthen mutant suppression, especially under strong stalling, because wildtype alleles can continue to amplify while 3′-proximal mutants are effectively diluted out before they elongate sufficiently to act as a template[23, 24, 35, 22].

Despite these limitations and potential extensions, our work shows that genomic architecture can act as an intrinsic bias for the fate of stored information. That is, by filtering which novelties are propagated versus rejected, the architecture effectively tunes the error threshold without explicit proofreading. The Virtual Circular Genome model was proposed[19, 20] to explain how primitive replication systems could avoid issues associated with the replication of long linear genomes such as strand separation and replication of terminal bases. Surprisingly, we have now shown that it also mitigates the unrelated problem of high intrinsic error rates, and thereby provides a means of avoiding the Eigen error catastrophe[4] and thus allowing for the emergence of Darwinian evolution.

Comparable architecture-level filters might be operating in extant biology, for example how RNA viruses with segmented genomes can undergo genetic reassortment[52]. Other non-standard genome architectures include those of multipartite bacteria (e.g., Borrelia burgdorferi[5, 53]), which partition essential genes across a chromosome plus several plasmids; ciliates package somatic information into vast numbers of gene-sized macronuclear chromosomes[7]; and trypanosomatids[8], which maintain kinetoplast DNA as a topologically linked network of maxi- and minicircles; and mitochondria and chloroplasts carry tens to thousands of genome copies per cell[6]. Even among close relatives, karyotypes differ sharply, e.g., Saccharomyces cerevisiae has 16 nuclear chromosomes whereas Schizosaccharomyces pombe has 3. In all these cases, information is distributed across many molecules with heterogeneous maintenance and inheritance, rather than on one continuous polymer. We lack a general framework for how such architectures influence mutation fate, dosage control, recombination, assortment noise, and long-term evolvability or why particular architectures arise and persist. While we do not address architectural features specific to these extant systems, our work highlights open questions about the advantages of distributing genomic information across many versus few molecules.

Acknowledgements.

Acknowledgments. We thank Irene Chen, Deborah Fygenson, Paul Higgs, Pepijn Moerman, Erik Winfree, and members of the Murugan and Szostak labs for helpful discussions. This work was supported in part by grants from the Sloan Foundation (19518) and the Gordon and Betty Moore Foundation (11479) to J.W.S. and A.M. J.W.S. is an Investigator of the Howard Hughes Medical Institute. A.M. acknowledges support from NIGMS of the National Institutes of Health under award number R35GM151211. M.J.F. is supported by the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, a Schmidt Sciences program. This work was supported by the National Science Foundation through the Center for Living Systems (grant no. 2317138) and by the University of Chicago’s Research Computing Center.

Data Availability. All code to run the simulations is publicly available[54]. Experimental data will be publicly available upon publication.