EnzymeRSI: Improving the Research Workflow, Not Just the Enzyme

A new preprint from the Genoria AI Team and colleagues proposes that enzyme-design systems should be able to revise not only their candidate lists, but also the way they conduct research. Called EnzymeRSI, the system separates changes to the research workflow from changes to the algorithms that predict and select variants. The distinction is useful: a misplaced workflow step calls for a different remedy than a predictor that cannot rank candidates reliably.

The authors report computational improvements and experimental decision-making across three enzymes. But their most ambitious vision—a system that repeatedly improves itself while commissioning and responding to laboratory experiments—has not yet been demonstrated end to end.

Two loops, different jobs

The outer loop changes the harness: prompts, tool access, pipelines and other arrangements around an AI model. It does not change the model’s weights. Proposed revisions are compared with the current version under matched conditions and can be adopted or rolled back.

The inner loop addresses a named methodological bottleneck, such as poor ranking, after engineering and workflow explanations have been considered. It can develop prediction and selection methods under controlled conditions. The authors also specify a cloud-laboratory route for returning measurements with quality-control information. A successful software receipt, they stress, is not by itself proof that a physical transfer or measurement occurred.

On five ProteinGym assays, outer-loop changes increased the average development-set Spearman correlation—a measure of ranking agreement—by 0.044. The CD19 binding endpoint reached 0.335, ranking first alongside the published models in the authors’ comparison. Rhodopsin expression also ranked first at 0.683, but its starting tool combination was already ahead; evolution contributed just 0.003. These were development-set results on assays used during evolution, not proof of transfer to new assays.

In the inner-loop stability task, the system reached a Pearson correlation of 0.849, compared with 0.834 for the best of ten reference frameworks and 0.840 for the unmodified QwenCode base. On a task predicting the ranking of later-round TadA variants, it reached a Spearman correlation of 0.307 versus 0.301 for QwenCode, using 1.13 rather than 2.24 active hours for 20 formal evaluations. These comparisons are informative but should not be treated as definitive framework rankings: the authors call for replication across independent evolution seeds, and reference frameworks did not all run identical numbers of native rounds.

What the enzyme tests reveal

The three laboratory cases examined different decisions, rather than a complete campaign driven by both loops.

For Renilla luciferase, EnzymeRSI froze 40 single-substitution nominations before the round’s measurements. Eight were components of a pre-specified variant; of the other 32, ten reached at least twice the same-plate wild-type integrated luminescence on retest. Yet candidate design rank correlated only weakly with retest rank: Spearman ρ was 0.175. The system assembled a useful pool but was much less successful at ordering it.

Measurement quality matters to that conclusion. Wild-type reference wells were variable, and sensitivity analysis showed that the twofold classification of some inference-only candidates depended on which reference wells were included. All three inference-only candidates remained above wild type under the tested removals, but only one remained above twofold throughout.

For APOBEC3A, the system applied a rule to measured data from 75 variants: limit conversion of unmodified cytosine in both sequence contexts, then maximize the weaker of the two 5-methylcytosine conversion rates. The preferred variant changed as the constraint changed. At a 50% cap on unmodified-cytosine conversion, Y130Q was selected; it converted 5-methylcytosine at 90.2% in one context but 46.5% in the weaker context. This was a recommendation from measured variants, not evidence that the same performance would hold under all conditions.

For KOD-E polymerase, the system analyzed measured variants and possible paths through them. One combination reached 31.1 times the parent’s relative apparent catalytic efficiency in the reported assay. The kinetic measure was obtained while DNA concentration varied and modified dATP was held fixed, so it should not be read as an absolute affinity for the modified nucleotide. The analysis also showed that a substitution’s effect could reverse in different sequence backgrounds: the best next step depends on both the existing variant and the objective.

The test still ahead

EnzymeRSI’s strongest contribution may be making failures actionable. The luciferase ranking gap, for example, is a plausible reason to invoke method development—but the preprint does not show that the inner loop has repaired it. Nor has the specified commission-and-feedback pathway yet run a multi-round physical campaign driven by both loops.

The next convincing demonstration would combine independent computational replication with that missing laboratory cycle: nominate, measure, diagnose, revise and repeat, while showing whether each revision actually improves decisions. Until then, the work is best read as a structured proposal backed by encouraging, bounded tests—not a demonstration of autonomous enzyme engineering at full scale.