{"id":"1b0cd484-1f52-48d0-8ece-66c618ac8f4e","arxiv_id":"2412.06115","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Protein evolution is best understood as a complex adaptive system, and this framing could guide new predictive models.","lead":"This perspective argues that protein evolution behaves like a complex system, with nonlinear dynamics, contingency, and emergent order. It suggests that viewing evolution this way could improve machine learning models and protein engineering.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central 'strange attractor' explanation is an unsupported analogy: UMAP clusters and phylogenetic branching are not shown to have the Lyapunov instability and fractal dimension that define strange attractors, and stochastic evolutionary contingency is not deterministic sensitivity to…","rationale":"The reader's verdict (CONDITIONAL) is appropriate. The paper is a perspective, so qualitative uses of terms can be acceptable, but the Summary makes a positive assertion ('embodies', 'explains') that goes beyond a heuristic. The weakest point is the strange-attractor claim because it is central to the paper's distinctive angle and is not backed by any dynamical evidence. My proposed check would decide whether real clusters in sequence space have the defining properties of strange attractors; I strongly suspect they do not, since a finite Markov chain has no strange attractor, but the test is what would settle it. This does not change the reader's conditional verdict: the paper should be accepted only if the authors soften or operationalize these claims. I also considered the glossary's definition of self-organization as 'without external factors' versus the text's reliance on selective pressure as an external force; that is a secondary inconsistency and could be fixed by clarifying that selection is treated as part of the coupled system. It does not displace the strange-attractor concern as the primary load-bearing issue.","tokens_in":16506,"tokens_out":7847,"duration_ms":79474,"concrete_test":"Use the CDH/SBP data behind Fig. 2d to make the claim testable: (1) define an explicit discrete-time mutation–selection map on that sequence space, with fitness from the experimentally characterized ancestral/extant variants; (2) estimate the largest Lyapunov exponent from ensembles of trajectories starting one substitution apart; (3) compute the correlation dimension of the high-fitness sequence cluster under a sequence metric. If the Lyapunov exponent is not positive, or the correlation dimension is integer (or the chain is finite and ergodic, so no strange attractor is possible), then the Summary's strange-attractor claim is unsupported and should be softened. A null-model UMAP with permuted functional labels would further show whether the Fig. 2d clusters are simply phylogenetic signals rather than attractor structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Most load-bearing is the Summary's claim that 'the presence of strange attractors, such as stable folds or functional states' demonstrates 'deep alignment between protein evolution and complex systems theory.' The evidence is §Strange attractors and fractal geometry: Fig. 2d shows a UMAP embedding of ancestral and extant CDH/SBP sequences, with functional clusters labeled 'strange attractors.' UMAP is a dimensionality-reduction tool; clusters in it reflect sequence relatedness and functional conservation, not an invariant set of a dynamical flow. No Lyapunov exponent, fractal dimension, or divergence-rate measurement is reported. The paper itself defines protein evolution as 'discrete' and 'indeterministic' (§Protein evolution is a complex system), while strange attractors are defined for deterministic systems. Additionally, 'sensitivity to initial conditions' is equated with evolutionary contingency (epistasis, entrenchment, historical accident), but contingency is stochastic path dependence, not the exponential divergence of nearby deterministic initial conditions that defines chaos. Thus the central explanatory claim is an unoperationalized metaphor: 'strange attractor' adds no testable content beyond saying that sequences converge to functional regions without repeating, which follows trivially from finite combinatorial space and purifying selection. The paper could be made defensible by reframing these as 'basins of attraction in a stochastic fitness landscape' and treating complex-systems terms as analogies rather than demonstrated properties.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This perspective argues that protein evolution should be understood as a complex system, exhibiting nonlinear dynamics, sensitivity to initial conditions, self-organization, and the emergence of order from chaos. It develops a series of analogies: evolutionary contingency as chaos-theoretic sensitivity to initial conditions, neutral drift as turbulence, fitness peaks and stable folds as attractors, UMAP sequence clusters as strange attractors, and phylogenetic trees as fractal objects. The paper then surveys machine-learning methods for modeling protein evolution and proposes future directions rooted in complex-systems tools such as bandit theory, information bottlenecks, and agent-based models. The central claim, restated in the Summary, is that the presence of strange attractors such as stable folds or functional states demonstrates a deep alignment between protein evolution and complex systems theory.","tokens_in":16782,"tokens_out":3773,"duration_ms":42868,"significance":"The paper is a synthesis rather than a new derivation, and its value lies in drawing together a broad and current literature on epistasis, entrenchment, contingency, de novo gene emergence, marginal stability, consensus design, and machine-learning models of protein evolution. It is commendably candid about the limitations of protein language models, including the observation that simple one-hot encodings can match or outperform LLM embeddings and that these models are best viewed as 'stochastic parrots.' If its proposed analogies were operationalized, the perspective could motivate quantitative tests using in silico evolution, ancestral sequence reconstruction, and high-throughput fitness measurements. However, the scientific significance of the central claim is currently limited because the most load-bearing concept, the 'strange attractor,' is used metaphorically rather than demonstrated, and no testable predictions follow from it beyond a general research agenda.","major_comments":[{"comment":"The identification of UMAP clusters as 'strange attractors' is unsupported and is load-bearing for the Summary's claim of 'deep alignment between protein evolution and complex systems theory.' UMAP is a dimensionality-reduction method; clusters in the embedding reflect sequence relatedness and functional conservation, not invariant sets of a dynamical flow. No Lyapunov exponents, fractal dimensions, or trajectory divergence rates are reported, so none of the defining properties of a strange attractor is established. Moreover, the paper explicitly describes protein evolution as 'discrete' and 'indeterministic' in an earlier section, whereas strange attractors are defined for deterministic dynamical systems. The authors should either provide a quantitative dynamical-systems analysis (e.g., measuring recurrence or divergence from reconstructed ancestral trajectories) or, more realistically for a perspective, reframe these objects as 'basins of attraction in a stochastic fitness landscape' and soften the Summary accordingly.","section":"Strange attractors and fractal geometry; Figure 2d; Summary"},{"comment":"The paper equates 'sensitivity to initial conditions' with evolutionary contingency, but these are different phenomena. In chaos theory, nearby deterministic initial states diverge exponentially under the same evolution rule; in protein evolution, divergent outcomes arise because stochastic mutation and epistasis produce different realized paths from similar or even identical starting sequences. The cited examples, such as hormone receptor evolution and directed-evolution reversibility, are cases of historical contingency and entrenchment, not demonstrations of deterministic chaos. A concrete discriminator would be replicate evolution experiments from identical starting sequences: chaotic determinism would require the same deterministic map with slightly different initial conditions, whereas the observed spread of outcomes is a distribution over stochastic trajectories. The authors should reframe this section as 'sensitivity to perturbations and historical contingency' and avoid implying that the chaos-theoretic definition has been satisfied.","section":"Initial conditions, contingency and directionality"},{"comment":"The 'turbulence' analogy is presented with the statement that neutral drift 'bears striking resemblance to turbulence,' but no quantitative aspect of fluid turbulence—such as an energy cascade, a Reynolds-number analog, or a defined disorder measure—is transferred to sequence space. The later discussion of laminar-to-turbulent transitions in protein disorder is also purely metaphorical. This would be acceptable as a heuristic if the paper consistently marked it as such, but the Summary elevates these analogies to evidence for 'deep alignment' with complex systems theory. The authors should either provide an operational definition of the proposed turbulence analog (for instance, in terms of variance in sequence entropies or flux across fitness thresholds) or explicitly downgrade these passages to intuitive illustrations.","section":"Turbulence, entropy and self-organization"}],"minor_comments":[{"comment":"The glossary entry for 'Analytical solution' is unusual: the example 'sequence after ten generations, given the initial sequence' describes a forward simulation or recurrence, not an analytical solution in the usual closed-form sense. Please clarify the intended distinction.","section":"Glossary"},{"comment":"The statement that phylogenetic trees 'mirror the fractal patterns observed in physical trees' is presented without any fractal-dimension analysis. Self-similar branching is a common property of random and neutral branching processes; calling it fractal requires a quantitative measure such as a scaling exponent.","section":"Strange attractors and fractal geometry"},{"comment":"The glossary defines a strange attractor as 'e.g. due to bifurcation through evolution.' Strange attractors arise through bifurcations in the parameters of a dynamical system, not 'through evolution'; this wording may confuse readers. Consider revising to distinguish parameter bifurcation from the paper's use of 'bifurcation' for mutational branching.","section":"Strange attractors and fractal geometry"},{"comment":"The passage on PINNs states that they are 'effective in modeling both chaotic systems, like the double pendulum, and systems with partial or noisy observations.' A double pendulum is chaotic but the cited PINN reference is specifically about PINNs failing to predict its chaotic motion; the paper should acknowledge this nuance or cite a more suitable example.","section":"Machine learning in the study of complex systems"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies noticeably on the authors' own prior work (e.g., refs 48, 56, 85, 106), but those citations are directly on-topic and do not by themselves justify concern. The larger issue is fit between the perspective genre and the strength of the claims: a perspective is allowed to be speculative, but the Summary states the strange-attractor claim as an established fact. If the authors are willing to reframe the central analogy as a testable hypothesis and add a short 'operationalization' paragraph describing what quantitative evidence would be required to confirm or refute it, the paper could be suitable for publication. As it stands, the paper overreaches in its most conspicuous claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a perspective, not a new result. It argues protein evolution has the hallmarks of complex systems—nonlinearity, sensitivity to initial conditions, self-organization, order from chaos—and uses that framing to survey ML approaches and suggest future tools like bandit theory. The best parts are the ML sections: the authors are honest about language models as 'stochastic parrots' and cite evidence that simple one-hot embeddings can match LLM encodings. That is a balanced, useful summary.\n\nThe soft spot is the strange attractor claim, and it's not minor. The Summary states that 'the presence of strange attractors, such as stable folds or functional states, highlights the deep alignment between protein evolution and complex systems theory.' But the evidence in Section \"Strange attractors and fractal geometry\" is a UMAP plot with clusters labeled as strange attractors. UMAP is dimensionality reduction; clusters there reflect sequence relatedness, not invariant sets of a dynamical flow. No Lyapunov exponents, no fractal dimension, no divergence rate are reported. The paper itself defines evolution as discrete and indeterministic, while strange attractors are defined for deterministic flows. Equating contingency with sensitivity to initial conditions conflates stochastic path dependence with deterministic chaos. That is a genuine inconsistency, and the claim as written is unsupported.\n\nThat said, this is fixable. The authors already describe fitness maxima and folds as attractors in a way that works as an analogy. If they reframe 'strange attractors' as 'basins of attraction in a stochastic fitness landscape,' or explicitly say they are using complex-systems language metaphorically, the perspective stands. The core idea—that protein evolution is a complex adaptive system whose trajectories converge on functional regions without exactly repeating—does not require the literal dynamical systems apparatus. So the central argument holds up; the execution overreaches. The self-citation is not a problem; the illustrative examples come from their own prior work, which is fine for a perspective.\n\nWho is this for? Researchers in protein evolution and ML who want a broad framing and a literature sweep. Not for people looking for new equations or data. It would make a reasonable reading-group piece to debate where analogies are helpful and where they mislead.\n\nMy recommendation: send it to peer review, but with a request for major revision focused on the strange attractor language. A good referee can help them say what they mean without the unsupported technical dressing. The paper is honest and well-written; it just needs to be disciplined.","headline":"A well-written perspective that oversells its central metaphor—the 'strange attractor' label needs to be softened, but the complex-systems framing and ML discussion are worth a serious look.","tokens_in":17310,"tokens_out":2956,"would_cite":false,"duration_ms":28745,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This perspective argues that protein evolution should be understood as a complex adaptive system, and that chaos-theory concepts—especially strange attractors—explain why evolutionary trajectories converge on recurrent folds and functions…","keywords":["protein evolution","complex systems","chaos theory","strange attractors","fitness landscapes","epistasis","protein language models","directed evolution"],"falsifier":"Take a fitness landscape measured by deep mutational scanning, evolve many populations from near-identical starting sequences under controlled selection, and ask whether outcomes diverge as predicted by sensitivity to initial conditions and whether the set of converged sequences has non-integer fractal dimension. If trajectories are largely reproducible from the starting sequence, or if converged sequence clusters have ordinary, non-fractal geometry, the strange-attractor description would be a metaphor rather than a mechanism.","tokens_in":16350,"feed_emoji":"🧬","tokens_out":4692,"duration_ms":47909,"temperature":0.7,"pith_summary":"This paper is a perspective, not a new experiment. It argues that protein evolution has the defining hallmarks of a complex system: nonlinear interactions, sensitivity to initial conditions, self-organization, and order emerging from chaos. The payoff would be a new modeling strategy: instead of trying to predict exact evolutionary trajectories, researchers should characterize the basins and attractors that constrain trajectories, using machine learning and in silico evolution. If the authors are right, protein engineering and design can be guided by the same tools used to understand turbulence and chaotic dynamics.","feed_headline":"Protein evolution is chaos with hidden order, review argues","feed_subtitle":"Mapping mutation and selection onto strange attractors and phase space explains why evolution converges without repeating.","key_machinery":"The central object is the strange attractor, a fractal subset of a dynamical system's phase space that nearby trajectories converge toward while remaining sensitive to perturbations. The paper maps protein sequence space to phase space, identifies stable folds and fitness maxima as attractors, and reads phylogenetic trees as fractal bifurcation patterns generated by repeated mutation-selection cycles. The strange attractor does the argument's main work: it explains why evolution appears to converge on the same functions and folds while never retracing the same sequence path. The treatment is analogical and illustrative; the paper does not compute fractal dimensions or Lyapunov exponents for any real protein system.","core_discovery":"The central claim is that protein evolution embodies the defining characteristics of complex systems. In this view, mutation and selection are coupled nonlinear forces that make outcomes exquisitely sensitive to starting sequences, which the paper identifies with evolutionary contingency; epistasis makes each mutation's effect depend on genetic background; and the interplay of drift and selection can be described as turbulence, with selection acting as an external force that pushes a disordered pool of sequences toward low-entropy states such as stable folds. The paper's distinctive proposal is that native folds and fitness maxima behave like strange attractors in sequence space: trajectories converge toward them but, because of sensitivity to initial conditions and the vastness of sequence space, never exactly repeat. This is presented as a conceptual framework for interpreting existing data—phylogenetic trees, enzyme evolution experiments, deep mutational scans, and protein language model behavior—rather than as a mathematical derivation.","pith_inferences":["Editorial extension: the paper uses 'strange attractor' as a metaphor, but the metaphor becomes a quantitative claim if future work estimates Lyapunov exponents or correlation dimensions from high-throughput evolutionary lineages, turning the analogy into a measurable dynamical property.","Editorial extension: if the complex-systems view is correct, protein language models may be implicitly learning attractor geometry; probing their latent spaces for funnel-like structure could predict evolvability better than current fitness benchmarks.","Editorial extension: the framework suggests that de novo gene birth and fold switching are not anomalies but expected turbulence-to-order transitions, which could guide experiments that deliberately search for weakly selective conditions favoring such transitions."],"forward_implications":["Protein evolution models should abandon deterministic trajectory prediction and instead characterize basins of attraction and the statistical behavior of trajectories within them.","Machine learning models trained on protein sequences should be evaluated not only on fitness prediction but on whether their latent spaces reproduce the attractor structure of real sequence space.","In silico directed evolution should be framed as an exploration-exploitation problem, using bandit-style algorithms to balance mutation toward known folds against sampling unexplored regions.","Consensus design and thermostability engineering are reinterpreted as applying selective order to counteract mutational turbulence, aligning sequences with a low-entropy equilibrium state.","Irreversibility and entrenchment imply that reverse evolution experiments will typically fail even when ancestral functions are selected for, because epistatic ratchets have closed the reverse paths."],"supporting_citations":[{"why":"Supplies the canonical example of sensitivity to initial conditions that the paper applies to evolutionary contingency.","marker":"[2]"},{"why":"Provides the working definition of a complex system whose characteristic features the paper maps onto protein evolution.","marker":"[16]"},{"why":"Gives the Hsp90 deep-sequencing evidence for entrenchment and irreversibility that supports the contingency and directionality claims.","marker":"[54]"},{"why":"Documents diminishing returns and tradeoffs in enzyme optimization, used to support the turbulence and stabilization analogy.","marker":"[60]"},{"why":"Defines strange attractors and fractal basin boundaries in nonlinear dynamics, the central concept transferred to sequence space.","marker":"[82]"},{"why":"Provides the ESM3 example used to argue that protein language models generate probable sequences rather than simulate evolutionary trajectories.","marker":"[102]"},{"why":"Introduces evolutionary velocity from protein language models, used as evidence that ML can map possible evolutionary trajectories.","marker":"[103]"},{"why":"Shows that ancestral-sequence-informed embeddings can outperform or match large language model encodings, supporting the paper's caution about what PLMs actually learn.","marker":"[106]"}],"fun_headline_variants":["Protein evolution: strange attractors in sequence space","Chaos with order: protein evolution as a complex system","Why proteins evolve to the same folds but never repeat paths","Protein evolution: nonlinear, self-organizing, and convergent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that mathematical ideas built for continuous dynamical systems—phase space, attractors, fractal geometry—transfer meaningfully to the discrete, finite, stochastic space of protein sequences, so that a cluster of similar sequences in a projection plot can be treated as a strange attractor.","fun_headline_variants_meta":{"raw":{"variants":["Protein evolution: strange attractors in sequence space","Chaos with order: protein evolution as a complex system","Why proteins evolve to the same folds but never repeat paths","Protein evolution: nonlinear, self-organizing, and convergent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2258,"prompt_tokens":840,"completion_tokens":1418,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":1353}},"tokens_in":456,"tokens_out":1418,"duration_ms":11927,"temperature":1.0,"reasoning_tokens":1353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:58:51.278146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fitness landscape measured by deep mutational scanning, evolve many populations from near-identical starting sequences under controlled selection, and ask whether outcomes diverge as predicted by sensitivity to initial conditions and whether the set of converged sequences has non-integer fractal dimension. If trajectories are largely reproducible from the starting sequence, or if converged sequence clusters have ordinary, non-fractal geometry, the strange-attractor description would be a metaphor rather than a mechanism.","supporting_citations":[],"review_version":1}