Artificial Intelligence · arXiv:2611.06321 v2 · submitted 4 Nov 2026
EnzymeRSI: A dual-loop recursive self-improvement system for enzyme design
The paper says
“Enzyme engineering usually optimises which candidates to assay within a fixed model and workflow; when the bottleneck is a mismatched method or a failing workflow, that route is hard to escape.”
more of the abstract
We present EnzymeRSI, a dual-loop system for recursive self-improvement (RSI). The outer loop evolves harness assets without changing model weights; the inner loop develops prediction and selection algorithms under controlled conditions; a cloud laboratory system is specified for closing the physical loop with quality-controlled feedback. The outer loop raised the mean development-set Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); of the two first places, only CD19 was produced by the outer loop itself. In ThermoMPNN self-improvement the inner loop reached a PCC of 0.849, and on TadA-Bench a Spearman correlation of 0.307 in 1.13 h. We examined the system's experimental decision-making in three enzyme systems. For the luciferase RLuc, 10 of the 32 nominated substitutions not pre-specified reached at least twice wild-type activity in a single round. For the cytidine deaminase A3A, a variant selected by a rule that caps loss of unmodified C raised 5mC conversion in the weaker CT context from 11.4% to 46.5% and reversed the substrate preference in both contexts. For the DNA polymerase KOD-E, 189 single-substitution variants defined a landscape on which different objective functions selected divergent next steps from the same parent; one assayed combination reached 31.1 times the parent on a modified substrate. By treating the organisation of research and the development of algorithms as separately scored objects of evolution, and linking a frozen decision record to a physical measurement, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 5 Nov 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Promising dual-loop RSI architecture for enzyme engineering, but the central claim of an autonomous evolutionary paradigm remains aspirational until multi-round closed-loop wet-lab feedback is demonstrated. This line is wrong
1 major objection · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the references say.
The paper says, at Abstract, §1, §3
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):
“Third, the three enzyme tests examine the system’s experimental decision-making rather than a complete dry-to-wet cycle driven by the two loops; the commission, execution and receipt path of Section 4.6 is specified but has not yet carried a multi-round campaign end to end, which we treat as a scope decision for this report rather than an unresolved capability.”
Pith says: Objection 1 of 1 · major
The paper claims that EnzymeRSI 'provides an evolutionary paradigm for autonomous enzyme engineering' (b44, b6). However, the three wet-lab tests (RLuc, A3A, KOD-E) are single-round experiments: the outer loop did not evolve in response to those measurements, the inner loop was not triggered by the RLuc ranking gap (acknowledged in b173), and the commission–execution–receipt path has not carried a multi-round campaign (acknowledged in b174).
more of objection 1
The outer-loop gain of 0.044 Spearman is measured on development sets only (§3, b172). To support the claim of autonomous enzyme engineering, the system must demonstrate that it can improve its own decision-making through multi-round closed-loop wet-lab feedback. The paper's own stated follow-up step (b176) confirms this gap. I recommend either (a) tempering the central claim to reflect the current level of validation (e.g., 'offers an architecture and partial validation toward autonomous enzyme engineering'), or (b) including preliminary evidence from a multi-round closed-loop campaign. Without such changes, the claim is unsupported by the presented evidence.
5 minor objections
The paper says, at Fig. 2 caption
Pith says: minor objection 1 of 5
The caption for Fig. 2 states 'the horizontal axis counts the framework’s own outer-loop rounds, each framework on its own scale'. This is appropriate for cumulative-best plotting, but the differing round definitions make direct visual comparison of trajectories difficult. It would be helpful to add a note about the number of formal evaluations per round for each framework, or to provide an additional panel with a common x-axis.
The paper says, at §2.2, paragraph 2
Pith says: minor objection 2 of 5
The discussion of QwenCode on TadA-Bench states that QwenCode 'migrated across more than ten model architectures, stopped at 0.301'. The phrase 'migrated across' is ambiguous; please clarify whether this refers to different model checkpoints, different architectures in the search space, or something else.
show 3 more
The paper says, at §3, b171–b174
Pith says: minor objection 3 of 5
The three boundaries are clearly and honestly stated. However, the final sentence of §3 (b44) returns to the strong claim of 'evolutionary paradigm for autonomous enzyme engineering' without restating these constraints. Consider adding a qualifying sentence at the end of §3 to align the conclusion with the acknowledged limitations.
The paper says, at Table 2 caption
Pith says: minor objection 4 of 5
The table note reads 'ΔPCC is computed from full-precision values, so for some frameworks it differs by one unit in the last place from the difference of the rounded figures shown.' Please verify that the difference is at most 0.0001 (one unit in the last decimal place given); the current phrasing could be interpreted as a difference of 1.0.
The paper says, at §4.6, World Loop
Pith says: minor objection 5 of 5
The World Loop description is technically detailed, but the sentence 'A step that returns success evidences a software event only, whereas the quantity actually transferred, the reaction conditions and the identity of the product still require the corresponding measurement and quality-control support' could be made clearer by separating the software-event statement from the measurement-verification statement into two sentences.
What would settle it
Execute one complete dry-to-wet cycle: (1) outer loop proposes a change to harness assets (e.g., modify candidate generation prompt), (2) candidate variants are nominated using the updated harness and sent to wet-lab, (3) physical measurements return as Evidence with QC status, (4) outer loop evaluates whether to adopt or rollback based on measured hit rate or rank correlation. Repeat for at least two cycles on a single enzyme target (e.g., RLuc). Report whether the system improves wet-lab performance relative to a fixed-harness baseline over cycles.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 5 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 1 major objection, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Read this paper for its architecture, not for its headline claim. The dual-loop design—separating harness evolution from algorithm evolution—is genuinely new for enzyme engineering. The authors integrate a cloud laboratory specification and show prospective wet-lab results on three enzymes. That combination is a real step beyond prior RSI work that only touched coding benchmarks.
But the central claim—that this provides an evolutionary paradigm for autonomous enzyme engineering—rests on a premise that hasn't been tested. The paper's own limitations section says it plainly: the commission-execution-receipt path has not yet completed a multi-round campaign. The wet-lab experiments are single-round. The RLuc ranking gap is identified as a trigger for the inner loop, but that loop hasn't been run against it. The outer-loop gain of 0.044 Spearman is on the development set, and only one of five first-place endpoints came from evolution itself.
To be clear, the paper is well-structured and unusually transparent about what it hasn't done. That honesty deserves respect. The architecture is worth understanding. The prospective wet-lab hits on RLuc, A3A, and KOD-E show that the system can nominate reasonable candidates. But the paradigm claim requires a closed loop where physical measurements change the harness and algorithms in subsequent rounds. That hasn't happened yet.
The stress-test concern holds up on reading. The paper says the multi-round campaign is a "scope decision for this report." That's fair as a limitation, but it means the main conclusion is aspirational rather than demonstrated. The authors also note that comparisons between frameworks are often built on single runs and different budgets—this applies to their own comparisons too.
Who this is for: anyone working on RSI for scientific discovery, especially in protein engineering. The dual-loop architecture is a design worth learning from.
Would I send this to peer review? Yes. It deserves a serious referee. The architecture is novel, the transparency is exemplary, and the limitations are clearly stated. But the reviewer should press for a multi-round demonstration before accepting the paradigm claim.
The strongest objection
The paper's central claim (b44, b6) is that treating research organization and algorithms as objects of evolution provides an evolutionary paradigm for autonomous enzyme engineering. For this to hold, the system must be able to improve its own decision-making in response to physical measurements across multiple rounds. The paper acknowledges (b174) that 'the commission, execution and receipt path … is specified but has not yet carried a multi-round campaign end to end.' The three wet-lab tests (RLuc, A3A, KOD-E) are single-round experiments: candidates were nominated or analyzed using existing or newly measured data, but the outer loop did not evolve the harness based on those results, and the inner loop was not triggered by the RLuc ranking gap (b173). The outer-loop gain of 0.044 is measured on development sets only, and only one of the five assays (CD19) shows a gain from evolution itself (b172). Without a multi-round closed-loop wet-lab demonstration, the system has not shown that it can autonomously improve enzyme engineering outcomes through recursive self-improvement. The paper presents a promising architecture and honest limitations, but the central claim is not yet fully demonstrated.
The reader identified the missing multi-round wet-lab loop as the weakest assumption (b172–b174), and the paper explicitly states it as a scope decision. My analysis confirms this is the single most load-bearing concern: the central claim of an autonomous evolutionary paradigm cannot be assessed until the system has been tested in a closed-loop setting where physical results drive evolution of the harness or algorithms. The reader's verdict of CONDITIONAL is appropriate given this gap; I recommend UNCHANGED because the condition is already stated and the paper acknowledges it. The concrete test I propose operationalizes the reader's first condition by specifying a multi-round wet-lab campaign. The paper's other limitations (development-set evaluation, single first place from evolution, marginal gains) are secondary to this core gap. No additional concern beyond the reader's identification was found.
The referee's report
If fully validated, the dual-loop architecture would represent an important step toward autonomous scientific discovery in protein engineering. The separation of harness evolution from algorithm evolution, the explicit triggering conditions for each loop, and the prospective wet-lab validation of candidate nomination and path selection are novel and potentially impactful. The computational benchmarks are thorough and the wet-lab experiments are competently executed. However, the significance of the central claim is currently limited by the absence of a multi-round closed-loop wet-lab demonstration, which the paper itself acknowledges has not yet been carried out (b174). The results as presented demonstrate a promising architecture and partial validation, but the full paradigm of autonomous recursive self-improvement in physical experiments remains to be shown.
Pith's reading of the paper (a machine's summary, not the authors')
The paper aims to establish that treating research organization—such as prompts, tool routing, and subagent allocation—and algorithm development as separate, verifiable objects of evolution yields concrete improvements in enzyme design. Current machine-learning-guided directed evolution optimizes which candidates to assay within a fixed method, but the method itself can become a bottleneck. EnzymeRSI introduces an outer loop that evolves harness assets without changing model weights, and an inner loop that develops prediction and selection algorithms on demand. In computational evaluations, the outer loop raised the mean Spearman correlation by 0.044 across five ProteinGym assays, and the inner loop achieved a Pearson correlation of 0.849 on ThermoMPNN. Prospective wet-lab tests yielded 10 new RLuc variants with at least twice wild-type activity, an A3A variant with 90.2% 5mC conversion, and a KOD-E variant with 31.1-fold relative catalytic efficiency. The authors acknowledge that the full dry-to-wet closed loop has not yet been run end-to-end and that the RLuc ranking gap remains to be closed by the inner loop.
What carries the argument. The dual-loop architecture: the outer loop performs version evolution over prompts, tool skill definitions, subagent configurations, and system prompts (harness assets) through iterative paired comparisons under matched conditions; the inner loop develops predictors, training recipes, search operators, and acquisition functions only after engineering and workflow causes have been excluded. Both loops connect to a cloud laboratory system via ProtoPilot (protocol orchestration) and Operon-S1 (fast experimental decisions), which execute commissions and return evidence with quality-control status and execution receipts. A version is adopted only after a paired comparison passes a pre-defined adoption rule (global, scoped, or rollback), and the interaction between loops is assessed through a four-cell paired evaluation design.
Core claim. EnzymeRSI demonstrates that separating the organization of research (the harness: prompts, pipelines, skills) from algorithmic method development, and evolving both in separate loops, can improve fitness prediction and experimental hit rates in enzyme engineering. The outer loop, which does not modify model weights, raised the development-set Spearman correlation on ProteinGym by 0.044 across five assays, with the CD19 binding ranking becoming first place through self-evolution. The inner loop reached a Pearson correlation of 0.849 on the ThermoMPNN stability task and a Spearman correlation of 0.307 on TadA-Bench. In wet-lab tests, the system nominated single-substitution candidates before measurement, dynamically recommended variants under changing constraints, and identified evolutionary paths under epistasis. However, the paper states that the full closed loop—where the two loops commission, receive, and respond to several rounds of physical measurement—has not yet been executed, and the RLuc ranking gap (Spearman 0.175) is a trigger condition that has not been closed.
The premise it rests on. The outer-loop gain of 0.044 Spearman is measured on the development set that the harness was permitted to use during evolution, and the full dry-to-wet closed loop has not yet been run end-to-end, so the claim of an autonomous evolutionary paradigm for enzyme engineering rests on an untested multi-round feedback cycle.
If the paper is right (Pith's inferences, not the authors')
If EnzymeRSI's approach is validated, enzyme engineering campaigns could systematically improve their own prediction strategies and workflow organization over time, potentially reducing the number of wet-lab rounds required to find active variants.
The separation of outer and inner loops provides a framework for attributing performance gains to either organizational changes or algorithmic improvements, enabling targeted investment in the more impactful lever.
The RLuc ranking gap (design rank vs. measured rank Spearman 0.175) shows that a high hit rate does not guarantee reliable candidate ranking; the system treats this as a trigger for the inner loop to develop better calibration.
The A3A case demonstrates that the optimal recommendation can switch sharply with the constraint on unmodified-C loss, implying that static scorers are insufficient and that selection rules must be recompiled dynamically against the objective.
The KOD-E path analysis reveals that different engineering objectives (highest efficiency vs. longest monotonic path) lead to divergent next-step recommendations, and that the outer loop must explicitly compile the researcher's objective into selection and stopping rules.
The central claim of an 'evolutionary paradigm for autonomous enzyme engineering' rests on the untested full closed loop; the reported dry-lab gains and prospective single-round wet-lab experiments show promising components but do not yet demonstrate multi-round autonomous improvement with physical feedback.
The outer-loop gain of 0.044 Spearman is measured on the development set and only one of five first places (CD19) arose from self-evolution; the expression first place was inherited from the starting tool combination, suggesting that the marginal benefit of harness evolution may be modest for tasks already well-served by the initial configuration.
The inner-loop improvement on ThermoMPNN (PCC 0.849) involved introducing ESM2 representations, while the closest reference framework (OpenRSI, PCC 0.834) used ensembling of fine-tuned members; the small advantage (0.015) may reflect representation engineering rather than a fundamentally new discovery method.
The wet-lab tests examine the system's experimental decision-making in single rounds rather than a multi-round campaign driven by both loops; until the interaction term (Section 4.9) is estimated on wet-lab endpoints, the claimed paradigm remains a proposal with partial evidence.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Enzyme engineering usually optimises which candidates to assay within a fixed model and workflow; when the bottleneck is a mismatched method or a failing workflow, that route is hard to escape. We present EnzymeRSI, a dual-loop system for recursive self-improvement (RSI). The outer loop evolves harness assets without changing model weights; the inner loop develops prediction and selection algorithms under controlled conditions; a cloud laboratory system is specified for closing the physical loop with quality-controlled feedback. The outer loop raised the mean development-set Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); of the two first places, only CD19 was produced by the outer loop itself. In ThermoMPNN self-improvement the inner loop reached a PCC of 0.849, and on TadA-Bench a Spearman correlation of 0.307 in 1.13 h. We examined the system's experimental decision-making in three enzyme systems. For the luciferase RLuc, 10 of the 32 nominated substitutions not pre-specified reached at least twice wild-type activity in a single round. For the cytidine deaminase A3A, a variant selected by a rule that caps loss of unmodified C raised 5mC conversion in the weaker CT context from 11.4% to 46.5% and reversed the substrate preference in both contexts. For the DNA polymerase KOD-E, 189 single-substitution variants defined a landscape on which different objective functions selected divergent next steps from the same parent; one assayed combination reached 31.1 times the parent on a modified substrate. By treating the organisation of research and the development of algorithms as separately scored objects of evolution, and linking a frozen decision record to a physical measurement, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-05
The authors' main claim, as the classifier picked it out 1.00
“By treating the organisation of the research workflow and the development of algorithms as a single falsifiable dry-to-wet loop, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering.”
How the main claim is stated: asserted 0.94
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.75 · human participants 0.04 · controlled comparison 0.75 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.60 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Improvement in enzyme engineering is usually pursued by optimising which candidates to assay within a fixed model and workflow; when the bottleneck arises from a mismatched method or a failing workflow, that route is difficult to escape. | background 0.87 | 0.05 | 0.04 | asserted 0.98 |
| We present EnzymeRSI, a dual-loop system for recursive self-improvement (RSI). | contribution 1.00 | 0.83 | 0.08 | asserted 0.99 |
| The outer loop evolves harness assets without changing model weights, the inner loop develops prediction and selection algorithms under controlled conditions, and a cloud laboratory system is specified to close the physical loop with quality-controlled feedback. | method 0.94 | 0.73 | 0.07 | asserted 0.99 |
| By updating its prediction strategy through iterative comparison, the outer loop raised the mean development-set Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); of the two first places, only CD19 was produced by the outer loop itself. | result 1.00 | 0.95 | 0.11 | asserted 1.00 |
| In the self-improvement of the ThermoMPNN algorithm the inner loop reached a PCC of 0.849, and on TadA-Bench it reached a Spearman correlation of 0.307 in 1.13 h. | result 1.00 | 0.90 | 0.05 | asserted 1.00 |
| In three prospective wet-lab tests of the system's experimental decision-making, 10 of the 32 candidates that had not been pre-specified for the luciferase RLuc showed at least twice the wild-type activity; a recommended variant of the cytidine deaminase A3A raised 5mC conversion to 90.2%; and path selection in evolution raised the catalytic efficiency of the DNA polymerase KOD-E on a modified substrate to 31.1 times that of the parent. | result 1.00 | 0.95 | 0.09 | asserted 1.00 |
| By treating the organisation of the research workflow and the development of algorithms as a single falsifiable dry-to-wet loop, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering. | interpretation 0.97 | 0.61 | 0.81 | asserted 0.94 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-05: 1 request, 6,089 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (8)
Figures are the authors', reproduced from the paper.
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment