← News
Science Breakthroughs Science Breakthroughs Knowledge

A new biological function can be one good mutation away—and hundreds of quiet ones away

A new biological function can be one good mutation away—and hundreds of quiet ones away
Boolean Network Model type
Edges Rewired Network change
Power-Law Growth Main finding

In the largest networks in this study—1,000 logic gates—evolving a new input-output function took roughly a thousand mutation attempts on average. That may sound like evolution has a long climb ahead. But the result is striking for what it is not: the search did not become exponentially harder as the network grew. Its cost rose approximately in proportion to network size. In this model, the reason is simple and biological in spirit: as a system gets larger, more mutations are possible, but most changes do not touch the machinery that determines what the system does.

The finding is not a forecast that a real organism needs 1,000 mutations to acquire a new trait. The paper studies an intentionally stripped-down world of logic gates, rewired connections, and a small population-genetic model. Its value is more basic: it offers a mechanism for how the architecture of a genotype–phenotype map—the relationship between molecular variants and their effects—can set the pace of evolutionary change (Farheen et al., 2026).

The Science

A mutation changes a genotype; selection acts on the resulting phenotype. That translation is not straightforward. A genotype might be a DNA sequence or, in this study, a network’s wiring diagram. A phenotype is an observable function—in the model, the rule that converts inputs into an output. Many different genotypes can produce the same phenotype, and a single mutation can either change function, leave it untouched, or make it worse.

The scale of the problem in nature is immense. A sequence of 1,000 DNA letters has roughly possible versions. Yet evolution does not inspect them all. It moves through neighboring versions by mutation, with selection influencing which variants persist. How those neighbors are arranged—whether useful variants are close, clustered, or separated by long stretches of no change—can matter as much as the total number of possibilities.

Farheen, Boroda, Levien, and Kondev use Boolean networks to make that arrangement concrete. Each network has two external binary inputs, a collection of NAND gates, and an output. NAND is a logic operation that returns false only when both inputs are true. Because NAND gates can be combined to implement any Boolean function, the network can compute any of the 16 possible mappings from two binary inputs to one binary output. Its genotype is the network’s set of connections; its phenotype is the function it computes.

The researchers define a mutation as a rewiring event: choose an existing connection, remove it, and replace it with a connection from another permitted node. Network size stays fixed. They exclude cycles, because a directed loop would make the network’s input-output computation ill-defined. Fitness is the fraction of four possible input combinations for which the network’s output matches a chosen target. For example, a network that gets two of four cases right has fitness 0.5.

The evolutionary process begins with a network that computes NAND and aims for AND. A mutation can be beneficial, neutral, or harmful. The study models a haploid Moran population in the strong-selection, weak-mutation limit: selection acts before another mutation arrives, and a population can be represented by one resident network at a time. Mutants with higher fitness are more likely to take over; neutral ones can sometimes fix by chance. The simulations count every mutation attempt, including those that fail to spread, because both successful and unsuccessful attempts contribute to evolutionary time.

The paper runs 500 simulations at each network size from 5 to 1,000 gates, using a population size of 10,000 for its main analysis. The authors also report results for a population of 10 and for other target functions in supplementary figures.

lays out the network and the mutation process;

presents the central scaling result; and

illustrates the proposed explanation in terms of which edges actually contribute to the output.

What They Found

The average number of mutation attempts needed to reach AND rose approximately linearly with the number of gates. A power-law fit gave an exponent of 0.96, close to the value of 1 expected for a linear relationship. The fit across all 12 network sizes had an amplitude of 9.8. The median showed the same basic scaling, so the result was not driven only by a few unusually long simulations.

Reported scaling exponents are close to linear

Power-law exponents reported in the paper for total mutation attempts or the rate of encountering beneficial mutations. The low-population exponent for beneficial-mutation probability is negative because that probability falls with network size.

Reported scaling exponents are close to linear
LabelValue
Total mutations, population 10,0000.96 power-law exponent
Total mutations, population 100.9 power-law exponent
Beneficial-mutation probability, population 10-0.87 power-law exponent

The total is mostly made up of mutations that do not fix in the population. The number of non-fixing mutations grew with network size, while the number of fixing mutations remained roughly independent of it. That difference makes sense in this model: fitness can take only five values—0, 0.25, 0.5, 0.75, and 1. Starting at zero and moving upward through those levels requires no more than four beneficial steps, regardless of whether the network has five gates or a thousand. The waiting between those rare steps is where size matters.

A compact phenotype space, despite a vast genotype space

Counts specified by the model: two external binary inputs permit 16 Boolean functions; fitness can take five values, so at most four beneficial fitness increases are needed from zero to one.

A compact phenotype space, despite a vast genotype space
LabelValue
Possible Boolean functions16 functions
Possible fitness values5 functions
Maximum beneficial increases from fitness 0 to 14 functions

As networks grew, non-fixing mutations were increasingly neutral. A neutral mutation changes the wiring without changing the phenotype. The paper identifies a major source: many connections are not part of any route from the external inputs to the output. Changing one of these “non-computing” edges cannot change what the network computes. It is a mutation in the genotype that leaves the phenotype alone.

The authors call a connection “computing” if it lies on a path from an input to the output. Their simulations found that the average number of computing edges stayed independent of network size. But the total number of network edges grows with the number of gates, and each edge can be rewired in many ways. So the fraction of mutation opportunities that touch a computing edge falls as the network expands.

That gives the authors’ explanation for the nearly linear waiting time. A typical network has on the order of possible one-step rewiring neighbors when it has gates. Only on the order of of those mutations arise by rewiring a computing edge. If a roughly constant fraction of those relevant changes improve fitness, then the chance that any random mutation is beneficial is proportional to . The expected wait for a beneficial mutation is therefore proportional to .

In the main simulations, the probability of choosing a computing edge is approximately proportional to , while the probability that such a change is beneficial does not vary with network size. Together, those observations connect the local layout of the genotype–phenotype map to the overall evolutionary search. The authors’ point is not merely that large networks have more possible mutations. It is that most of those additional opportunities are functionally irrelevant.

At population size 10, the reported scaling remained roughly linear: the exponent for total mutation attempts was 0.9. The study also reports that the probability of a beneficial mutation in that condition scaled approximately as . At this smaller population size, neutral mutations are more likely to fix than in a population of 10,000, so the mix of fixing and non-fixing attempts differs. The supplied paper text cuts off before completing its discussion of this consequence, so the detailed implications for the small-population case cannot be fully assessed here.

Why This Changes Things

The result pushes against two simple intuitions. One says that bigger systems should evolve new functions faster because they offer more possible routes. Another says that a larger genotype makes the right combination exponentially harder to find. In this model, neither intuition wins outright. Bigger networks offer more mutations, but the number of functionally relevant connections does not grow at the same pace. The net effect is a steady, roughly linear increase in waiting time.

That is a useful way to think about biological complexity. A large genome, protein, or cellular network contains many components, but only some may strongly influence a particular trait under a particular environment. A mutation in a regulatory element can alter a key output; a mutation elsewhere may have no detectable effect. The “effective” search space depends not just on how many changes are possible, but on how phenotype-relevant changes are distributed among them.

The framework also links to the concept of evolvability: how readily a system can generate heritable variation that selection can use. A system can be robust—many mutations leave its current function intact—and still be evolvable if some accessible mutations lead toward new functions. In the Boolean model, neutral rewiring creates many function-preserving steps, but those steps mostly occur off the output-relevant paths. That architecture imposes a predictable search cost.

This matters for efforts such as directed evolution, where scientists repeatedly mutate and select proteins or other molecules to obtain a desired property. It also matters for thinking about pathogens and emerging threats. The model does not predict when a virus will evolve a particular trait; real evolution depends on mutation rates, population sizes, environments, epistasis (when the effect of one mutation depends on others), and the biology of the organism. But the study underscores why knowing the map from mutations to function can improve forecasts: the same number of mutations can have very different consequences in systems with different local organization.

The constraints are substantial. Two inputs and 16 phenotypes make this a tiny functional world compared with a cell, where outputs interact and environments shift. The model keeps network size fixed, uses a specific rewiring rule, and starts from a particular NAND network before targeting AND. It excludes cycles and uses a strong-selection, weak-mutation regime. The paper’s conclusions therefore apply most directly to this formal model, not as universal laws of molecular evolution.

Nor should the fitted exponent be mistaken for a universal biological constant. The central theoretical argument depends on a particular structure: the number of output-relevant edges remains roughly fixed while the total number of mutation opportunities grows. In a real network, the active part might grow with system size, mutations might alter several components at once, or selection might favor neutral paths that reshape the network in useful ways. Any of those could change the scaling.

What’s Next

The next test is to ask which parts of this mechanism survive when the model becomes more biological. Researchers could vary the number of inputs, use different network architectures, let the number of functionally relevant connections scale with network size, or allow multiple mutations and environmental changes to interact. They could also compare the predicted patterns with measured mutational data from genes, proteins, and engineered regulatory systems.

A particularly important question is whether real genotype–phenotype maps share the model’s imbalance between total and function-relevant mutation opportunities. If they do, the number of components in a system may be less informative than how many of those components can affect a trait. If they do not, the mismatch would identify exactly where richer biological mechanisms—neutral networks, modularity, or context-dependent effects—alter evolutionary search.

For now, the paper offers a precise conceptual takeaway rather than a timetable for evolution: new functions can be delayed not because useful mutations are vanishingly rare in absolute terms, but because a growing share of possible changes never reaches the function at all. Mapping that geometry could help scientists design more efficient laboratory evolution experiments and make better-grounded assessments of how biological systems adapt. The work turns a broad question—how many mutations to a new function?—into a sharper one: among all the changes a system can make, how many actually touch the computation that matters?