How Evolution Turns Altruists Into Parasites' Worst Nightmare

Imagine a world where kindness inevitably becomes corruption — where species that once helped each other reproduce inevitably evolve to serve themselves first. Now imagine that this selfish turn actually makes the entire community stronger. This is the paradox at the heart of a new mathematical model of evolution, and it may explain one of biology's most enduring puzzles: why cooperation exists at all, and how life protects itself from parasites that should, by all rights, destroy it.

The model, developed by mathematicians Alexander Bratus, Sergey Drozhzhin, and Tatiana Yakushkina at Moscow State University and the Yerevan Physics Institute, treats evolution not as a fixed landscape of fitness peaks and valleys — the classic metaphor of evolution as a mountain climber finding the highest point — but as a landscape that itself evolves. The fitnesses of species, the interactions between them, the very rules of the game: all of these shift over evolutionary time, and the system climbs toward better configurations not despite this dynamism but through it.

The finding that emerges from their mathematical analysis is striking: across four different biological systems — a molecular hypercycle, a double hypercycle, an anthill colony, and a network of RNA molecules — evolution follows an identical three-phase script. First, the system grows fitter without changing its basic structure: species maintain equal abundance, and all replication is purely altruistic, each species helping its neighbor reproduce. Then, suddenly, the equilibrium shifts. Dominant species emerge. The altruism begins to mix with selfishness — species develop the ability to catalyze their own reproduction alongside their neighbors'. Finally, the system stabilizes, settling into a configuration that resists invasion by parasites. The whole process takes hundreds of evolutionary steps, but it happens in the same way every time (Bratus et al., 2026).

This universality is the real surprise. Mathematical biology has long studied evolution as a process of optimization, but most models assume that the goal — the fitness landscape — is static. Real evolution doesn't work that way. The landscape shifts as the species themselves change it. What Bratus and colleagues show is that even when both the species and their fitness landscape evolve together, the outcome follows a predictable path. Fitness always increases. Cooperation always shifts toward a mixture of self-interest and mutual aid. And every system, regardless of its initial structure, acquires resistance to parasitic species that would have destroyed it in its original form.

The Science

To understand what this paper does, you first need to understand what a replicator is. In the mathematical framework that underlies much of evolutionary biology, a replicator is any entity capable of three things: copying itself, passing on hereditary variation, and competing for limited resources. This definition spans an enormous range of biological phenomena — from individual genes in a population to species in an ecosystem to molecules in a primordial soup. The replicator equation, which describes how the frequencies of competing species change over time, was formalized in the 1970s and has since become a cornerstone of theoretical biology (Hofbauer and Sigmund, 1998).

The equation governing an n-species replicator system looks like this:

$$\frac{d{u}_i}{dt}=u_i\left[(\mathbf{Au}{)}_i-f(\mathbf{u})\right]$$

Don't let the notation intimidate you. The variable $u_i$ represents the frequency of species $i$ in the population — a number between 0 and 1 that tells you what fraction of the community belongs to that species. The matrix $\mathbf{A}$ encodes the fitness landscape: the entry $a_{ij}$ tells you how much species $j$ contributes to species $i$'s fitness. If $a_{ij}$ is positive, species $j$ helps species $i$ reproduce. If it's negative, species $j$ competes with or harms species $i$. The term $f(\mathbf{u})$ is the mean fitness of the whole population, calculated as the average contribution across all species.

This equation has a simple intuition. A species grows when its individual fitness — its ability to reproduce — exceeds the community average. It shrinks when it falls below. The term $(\mathbf{A}\mathbf{u})_i$ captures everything that affects species $i$'s fitness: the support it gets from neighbors, the competition it faces, its own self-replication capacity.

The classic example is the hypercycle, a closed loop of species where each one catalyzes the next in a ring. Species 1 helps species 2, species 2 helps species 3, and species $n$ helps species 1. No species helps itself directly. This is pure altruism: every act of catalysis benefits a neighbor, not the catalyst. And yet, mathematically, such a system can persist. Hofbauer and Sigmund proved in the 1990s that hypercycles of order $n \geq 2$ are permanent — they maintain all species indefinitely — because the cyclic catalysis creates a mutual dependence that stabilizes the whole structure (Hofbauer and Sigmund, 1998).

But here's the problem that has haunted evolutionary theory for decades. If hypercycles require pure altruism to function, why would natural selection, which favors self-interest, ever permit them to exist? The mathematics says they're stable, but evolution is supposed to favor cheaters — parasites that take the benefits of the system without contributing to it. Classical quasispecies theory, developed by Manfred Eigen in the 1970s, showed that at sufficiently high mutation rates, a master sequence population undergoes "error catastrophe": the information content collapses because too many offspring are corrupted copies. Parasites should do the same thing to cooperative systems.

Bratus and colleagues approach this puzzle from an unusual angle. Rather than asking what happens to a fixed fitness landscape over ecological time, they ask what happens when the landscape itself evolves over evolutionary time. They introduce a new parameter $\tau$ — evolutionary time — that operates on a much slower scale than the standard ecological time $t$. While species frequencies change rapidly as the system approaches equilibrium, the fitness landscape matrix $\mathbf{A}(\tau)$ changes slowly, almost imperceptibly.

This two-timescale separation is justified mathematically by Tikhonov's theorem, a result from the theory of singular perturbation that shows how systems with very different timescales can be analyzed by treating the fast dynamics as instantly reaching equilibrium while the slow dynamics plod along. The theorem guarantees that the long-run behavior of the coupled system can be understood by focusing on the equilibrium of the fast dynamics for each fixed value of the slow parameter — in other words, by studying how the steady-state mean fitness changes as the fitness landscape itself shifts.

The key question becomes: over the admissible set of all possible fitness landscapes — all matrices $\mathbf{A}(\tau)$ whose entries don't grow unboundedly — does evolution drive the system toward a maximum of the mean fitness? And if so, what does that maximum look like?

To answer this, the researchers define the admissible set $\mathcal{M}$ as all matrices whose squared entries sum to at most some constant $M$:

$$\mathcal{M}=\left\{\mathbf{A}(\tau )=(a_{ij}{)}_{i,j=1}^n:\sum _{i,j=1}^n{a}_{ij}^2(\tau )\le M=\text{const}>0\right\}$$

This constraint keeps the fitness landscape bounded — no species can become infinitely beneficial or harmful. It's a biologically reasonable assumption: in any real ecosystem, there are limits to how strong interactions can be.

The convexity of this set is crucial. Two matrices in $\mathcal{M}$ can be averaged together, and the result stays in $\mathcal{M}$. This property means the space of possible fitness landscapes is smooth and well-behaved — there are no holes or jagged edges where optimization might get stuck. Evolution can search this space systematically, and any local optimum is also a global optimum.

The mathematical heart of the paper is Theorem 2.2, which derives a fitness variation formula. If the equilibrium frequencies $\bar{\mathbf{u}}(\tau)$ and the fitness landscape matrix $\mathbf{A}(\tau)$ both vary smoothly with evolutionary time, then the rate of change of mean fitness is:

$$\frac{d\bar{f}}{d\tau }=\delta \bar{f}(\tau )=(\delta \mathbf{A}(\tau )\bar{\mathbf{u}},\bar{\mathbf{v}})\cdot \bar{f}(\tau )$$

The vector $\bar{\mathbf{v}}(\tau)$ is defined by the adjoint equation $\mathbf{A}^T(\tau)\bar{\mathbf{v}}(\tau) = \mathbf{I}$, where $\mathbf{I}$ is the vector of all ones. In practical terms, this formula says that the change in mean fitness at any moment depends on how the fitness landscape is changing — specifically, on the correlation between the changes in interaction strengths and the current equilibrium structure of the population.

The researchers show that necessary conditions for the mean fitness to reach a local maximum are given by:

$${\bar{u}}_i({\tau }^{\ast }){\bar{v}}_j({\tau }^{\ast })=\mu \cdot a_{ij}({\tau }^{\ast })$$

where $\mu$ is a constant. This condition relates the equilibrium abundances to the fitness interactions in a specific way. When it's satisfied, the system has found a configuration where further evolutionary change cannot increase fitness.

The sufficiency conditions — ensuring that this critical point is actually a maximum rather than a minimum or saddle — involve second-order terms that depend on the curvature of both the fitness landscape and the equilibrium response. The researchers prove that these conditions reduce to a linear programming problem at each step: given the current state, find the direction of change in $\mathbf{A}$ that maximizes the fitness increase, subject to the constraints that keep the system in the admissible set and maintain non-degeneracy.

What They Found

The most striking results come from the numerical simulations, where the researchers apply their evolutionary algorithm to four canonical replicator systems: the hypercycle, the bi-hypercycle, the anthill system, and the RNA molecule network.

The hypercycle results are the most thoroughly analyzed. Consider a fifth-order hypercycle — five species arranged in a ring, each catalyzing only its cyclic neighbor. No species helps itself. This is the purest form of altruistic replication: every benefit flows outward to the next species, never inward.

When the evolutionary algorithm runs on this system, something remarkable happens. For the first 125 steps of evolutionary time, nothing visible changes in the equilibrium frequencies: each species maintains an equal share of the population, $u_i = 1/n$. The mean fitness, however, rises monotonically throughout this phase. The system is becoming more efficient even while its structure appears frozen.

This is the first phase of the universal pattern. Call it the altruistic growth phase. The fitness increases because the matrix $\mathbf{A}$ is being refined — interaction coefficients are being adjusted to extract more benefit from the existing altruistic structure — but the equilibrium hasn't yet shifted. Species remain equal because no species has yet gained a competitive edge.

Then, around step 125 for the ninth-order system, the equilibrium suddenly bifurcates. The species stop being equal. Some begin to dominate while others decline. This is the second phase: dominant species emergence. The altruistic structure that maintained equality begins to crack.

But here's the twist: as dominant species emerge, they don't simply outcompete the others. Instead, the system evolves new interactions. The purely altruistic ring develops reverse connections — species begin to catalyze not just their forward neighbor but also their backward neighbor. More importantly, each species acquires autocatalytic self-replication links: the diagonal entries of the fitness matrix, which were zero in the original system, become positive. Pure altruism becomes mixed altruism and selfishness.

The interaction graph transforms from a simple ring to a more complex network. Each node still connects to its neighbors, but now each node also has a direct connection to itself. The system is learning to help itself.

The contrast between these two figures is striking. Figure 4 shows the original hypercycle: a clean ring of arrows, each pointing from a catalyst to the species it helps, with no arrows pointing back and no arrows looping to the self. Figure 5 shows the evolved system: the original arrows remain, but now there are reverse arrows and self-loops. The altruistic hypercycle has become a selfish one.

And yet, paradoxically, the system is now more robust. This is the third phase: stabilization. The evolved system has crossed a threshold analogous to the error catastrophe boundary in quasispecies theory, but in reverse. Instead of information collapsing due to too much mutation, the system has achieved a configuration that resists parasitic invasion.

The researchers test this by introducing parasites into both the original and evolved systems. A parasite in this context is a species that benefits from the catalytic interactions without contributing to them — it rides the network for free. In the original fifth-order hypercycle, a parasite quickly drives the system to extinction. The frequencies of the original species collapse, and the parasite takes over. This is the expected outcome: selfish exploitation destroys cooperative structures.

But in the evolved system, something different happens. When a parasite is introduced to the fifth-order hypercycle after 200 evolutionary steps, the system absorbs the shock and recovers. The parasite cannot establish. When the same parasite is introduced at step 250, later in the evolutionary trajectory, the resistance is even stronger. The system doesn't even show a visible perturbation in its dynamics.

These figures show the difference starkly. In panel (a) of Figure 6, the original hypercycle crumbles when a parasite appears: the frequencies of the original species (shown in color) collapse to near zero within 50 time units, while the parasite (in black) ascends to dominance. In panel (b), the evolved hypercycle barely notices the parasite: all species maintain their equilibrium frequencies, and the parasite dies out.

The same pattern holds for higher-order systems. When two parasites are introduced to a ninth-order hypercycle, the original system collapses within 100 time units. The evolved system, obtained after 200 evolutionary steps, again resists. The two parasites cannot establish in the adapted community.

The three-phase pattern is clearly visible in the equilibrium dynamics. During the first phase, species stay equal while fitness rises. During the second phase, species differentiate while fitness continues to rise and the interaction structure transforms. During the third phase, the system stabilizes at a fitness maximum that is resistant to invasion.

Three-Phase Evolutionary Pattern in a Ninth-Order Hypercycle

Label	Value
Phase 1: Altruistic Growth	125
Phase 2: Dominant Species Emergence	125
Phase 3: Stabilization	100

Data derived from Figure 2: Equilibrium coordinates remain equal (uᵢ = 1/9) for ~125 steps, then split. Total evolutionary trajectory spans ~350 steps.

Monotonic Increase in Mean Fitness During Evolution

Label	Value
τ = 0	0.2
τ = 50	0.3
τ = 125	0.45
τ = 200	0.58
τ = 275	0.68
τ = 350	0.72

Data derived from Figure 3: Mean fitness increases monotonically throughout evolution. Values are approximate representative points from the monotonically increasing curve.

Figure 2 (the data underlying chart 0) shows how the equilibrium coordinates $\bar{u}_i$ evolve for a ninth-order system. For the first 125 steps, all nine lines overlap at $u_i = 1/9$ — perfect equality. Then they split: some species rise, others fall. The graph doesn't show the further stabilization that comes later, but the implication is clear: after the splitting phase, the system would converge to a new set of values, with dominant species holding higher frequencies and subordinate species holding lower ones.

Figure 3 (the data underlying chart 1) shows the corresponding mean fitness trajectory. The fitness increases monotonically throughout — it never decreases, never plateaus prematurely. This is the mathematical expression of the evolutionary optimization: the system is always finding configurations with higher mean fitness, and it never gets stuck in local optima.

The bi-hypercycle, anthill, and RNA network systems show the same three-phase pattern, though with different details. The anthill system, which models a colony with a dominant queen species catalyzing a ring of workers, transforms its fitness landscape in the same way: the queen's dominance eventually gives way to a more distributed structure where workers develop self-catalytic capacity. The RNA network, which models six RNA molecules in two interlocking hypercycles, undergoes the same transition from pure altruism to mixed altruism-selfishness.

The universality of this pattern is the paper's central empirical finding. Whatever the initial structure — pure altruism, queen-dominance, dual hypercycles — the evolutionary algorithm drives the system through the same three phases to the same kind of endpoint: a fitness maximum with mixed interactions and parasite resistance.

But the researchers go further. They prove that this outcome depends on a crucial assumption: non-degeneracy. The systems they're studying are permanent — all species coexist in the long run. If you relax the non-degeneracy constraint, allowing the equilibrium to touch the boundary of the species space, the dynamics change dramatically.

Without non-degeneracy, evolution no longer leads to fitness maximization. Instead, it leads to sequential species annihilation. One by one, species go extinct, each extinction reducing the dimensionality of the system and increasing the mean fitness of the survivors. The researchers prove a spectral lower bound on how much fitness increases with each dimension reduction: the fitness gain from eliminating one species is at least proportional to some eigenvalue of the interaction matrix. But the process is destructive, not constructive. The system doesn't find a cooperative optimum; it finds a selfish one by eliminating the competition.

This contrast — cooperative optimization with non-degeneracy versus destructive optimization without it — is one of the paper's deepest results. It suggests that the stability of cooperative systems is not a given but a condition that must be maintained. When species can coexist, evolution climbs toward better configurations while preserving the community. When coexistence breaks down, evolution proceeds by elimination.

Why This Changes Things

The standard view of evolution, even in sophisticated mathematical treatments, treats fitness landscapes as fixed or slowly drifting. A species climbs a fitness peak, but the peak itself doesn't move. This is the framework that underlies much of evolutionary game theory, population genetics, and ecological modeling. It's a useful simplification, but it's not how evolution actually works.

In reality, species modify their environments. They change the fitness landscapes they inhabit. The rise of photosynthesizing organisms transformed Earth's atmosphere, creating new niches and closing others. The emergence of predators reshaped prey populations. The evolution of the immune system changed the selective pressures on pathogens. The landscape moves because the climbers are terraforming the mountain as they climb.

Bratus and colleagues offer a mathematical framework for thinking about this coupled evolution of species and fitness. Their key insight is that when the fitness landscape evolves on a slower timescale than the species dynamics — a separation that seems biologically plausible for most evolutionary processes — the adaptation problem reduces to a sequence of maximizations. At each instant, the system behaves as if it were at equilibrium, and the equilibrium itself shifts slowly toward higher fitness.

This is not the hill-climbing of classical optimization. It's something more subtle: a co-evolutionary process where the goalposts move, but only slowly, and where the system tracks the moving target by always staying near the current equilibrium. The mathematics shows that such a process can converge to a global optimum even in a non-convex landscape, provided the landscape evolves in the right way.

The universality of the three-phase pattern is the most practically significant finding. It suggests that evolution doesn't explore the space of possible interactions randomly. It follows a deterministic path: first optimize within the existing structure, then restructure, then stabilize. This path leads inevitably from pure altruism to mixed altruism-selfishness, and from vulnerability to resistance.

The parasite resistance result is the most immediately striking implication. The original hypercycle, the anthill, the RNA network — all are vulnerable to parasitic exploitation in their initial forms. But evolution transforms them into systems that resist parasitic invasion without any external pressure toward resistance. The immunity emerges as a byproduct of the fitness maximization process.

This has echoes in several biological debates. The "Red Queen" hypothesis suggests that coevolution between hosts and parasites drives continual evolutionary change: each adaptation by the host is matched by a counter-adaptation by the parasite. But Bratus and colleagues' model suggests that some systems achieve a kind of intrinsic resistance — not by special-purpose immune mechanisms, but by the general logic of fitness optimization. The cooperators that survive aren't those that fight hardest; they're those that evolve the right network structure.

The distinction between degenerate and non-degenerate systems is also significant. Permanence — the condition that all species coexist — is not just a stability property; it's a precondition for cooperative optimization. When species can go extinct, the system doesn't find cooperative maxima; it finds selfish ones by eliminating the competition. This suggests that the preservation of biodiversity isn't just a passive side effect of ecological balance. It's an active enabler of cooperative evolution.

Fisher's fundamental theorem, one of the oldest results in mathematical biology, states that "the rate of increase in fitness of any organism at any time is equal to its genetic variance in fitness at the time." This theorem has been debated, qualified, and reinterpreted for nearly a century. The researchers show that Fisher's theorem holds only for symmetric fitness matrices — cases where the benefit from $j$ to $i$ equals the benefit from $i$ to $j$. In asymmetric systems like the hypercycle, where species help their neighbors but don't receive help back, the theorem breaks down. The new framework rehabilitates Fisher's intuition in a different form: fitness increases along the evolutionary trajectory, but the increase is driven not by variance at a fixed time but by the change in the fitness landscape over evolutionary time.

What's Next

The paper leaves several questions open, both mathematical and biological.

On the mathematical side, the researchers have shown that the fitness maximization problem reduces to linear programming at each step, but they haven't characterized the global convergence properties in full detail. The convexity of the admissible set ensures that local optima are global, but the path to the optimum may be long and winding. The rate of convergence — how many evolutionary steps are needed to approach the optimum — is not analyzed.

The sufficient conditions for a maximum, derived in Theorem 2.2 and its corollaries, are technically complex. They involve second-order terms that depend on the curvature of both the fitness landscape and the equilibrium response. In practice, verifying these conditions for a given system requires numerical computation. An analytical characterization of the stabilizing phase — the conditions under which the system settles into an error-catastrophe-equivalent configuration — would be valuable.

On the biological side, the four systems studied are highly abstract. The hypercycle, the anthill system, and the RNA network are mathematical models, not direct representations of any real biological community. The researchers note that the hypercycle has been studied as a model for prebiotic molecular evolution — the idea that before DNA and proteins, there may have been a world of self-reinforcing molecular networks. The RNA network is explicitly a model of early life, based on the work of Spiegelman and others on RNA replication. But the jump from these abstractions to real ecosystems is not straightforward.

The assumption of two-timescale separation — fast ecological dynamics, slow evolutionary dynamics — is standard but not universal. Some evolutionary processes, like viral adaptation to new hosts, happen on the same timescale as ecological dynamics. The framework would need to be extended to handle coupled fast-slow systems where the timescales are not widely separated.

The resistance to parasites is shown for specific parasite species introduced at specific times. Whether the resistance extends to arbitrary parasites — whether the evolved system is globally resistant or just resistant to the particular invaders tested — is not proven. The mathematics suggests that the stabilized configuration is an attractor for the evolutionary dynamics, but whether it repels all possible parasites or just a subset is a question for future work.

Perhaps the deepest open question is about the origin of the non-degeneracy constraint itself. The paper assumes that the systems it studies are permanent — that all species coexist in the long run. But in nature, extinction is common. The transition from non-degenerate to degenerate dynamics — from cooperative optimization to sequential annihilation — is not modeled. What starts the extinction cascade? What prevents it in the first place? The contrast between the two outcomes is stark, but the boundary between them is not characterized.

These are not weaknesses in the paper so much as indicators of its depth. The mathematics opens more questions than it closes. The three-phase pattern, the parasite resistance, the altruism-to-selfishness transition, the contrast between cooperative and destructive optimization — each of these findings suggests a research program that could occupy theoretical biologists for years.

The practical implications, while speculative, are suggestive. If evolution inevitably transforms cooperative systems into resistant ones, this may explain why many biological communities appear to be in a state of stable equilibrium rather than constant flux. The hypercycle doesn't keep evolving forever; it reaches a fitness maximum and stays there. Real ecosystems may similarly reach quasi-stable configurations that persist until environmental change forces new adaptation.

The result also illuminates the logic of cooperation. Altruism — helping others at cost to yourself — is paradoxical from a selfish gene perspective. Why would natural selection permit it? The traditional answers involve kin selection, reciprocal altruism, group selection, and spatial structure. Bratus and colleagues add a new answer: altruism may be an evolutionary precursor, a stage that systems pass through on the way to a more robust configuration. Pure altruism is unstable not because it gets exploited but because it gets transformed. The altruistic hypercycle evolves into a mixed system that retains the cooperative structure while adding selfish self-catalysis. Cooperation doesn't survive despite natural selection; it evolves into something that natural selection can sustain.

This is, in the end, a story about optimization and resilience. The fitness landscape evolves, species adapt, and the system climbs toward higher mean fitness. Along the way, it passes through phases of pure cooperation, competitive differentiation, and stable equilibrium. It emerges not as a utopia of perfect altruism but as a mixed economy of self-help and mutual aid — and this mixed economy is more resilient than either extreme. The mathematics doesn't prescribe morality, but it does suggest a logic: cooperate first, then learn to help yourself, and you'll be resistant to those who would exploit you.

The paper is technical, but its implications are broad. It suggests that evolution is not the random walk of chance and necessity that Stephen Jay Gould imagined, nor the greedy optimization of narrow self-interest. It's something more structured: a deterministic process following a universal three-phase script, driven by fitness maximization on a slowly evolving landscape. The script may not be universal in all biological systems — the abstraction is significant, and real ecosystems have complications the model doesn't capture. But the core insight — that fitness landscapes evolve and systems optimize through coevolutionary dynamics — seems robust. The mountain moves, and the climbers follow, and the path they take is not random but follows a pattern that mathematics can describe.