Enzyme–reaction pairing moves beyond sequence similarity: RIVER pools evidence with multiple agents and places all four target enzymes near the top of 667 candidates
Synopsis
RIVER is a training-free multi-agent retrieval framework that organizes large language models, biochemical computational tools and external knowledge resources into traceable evidence chains for bidirectional enzyme-to-reaction and reaction-to-enzyme retrieval. In four experimentally supported enzyme–reaction cases drawn from an E. coli metabolomics study, it ranked among 667 candidate proteins with a mean reciprocal rank of 0.833 using Qwen3.6-35B-A3B and 0.750 using GPT-5.6 Luna, above the general-purpose coding-agent combinations and domain methods listed in the comparison table. On public benchmarks it reached Top-1 query success rates of 82.3% on Enzyme-405 and 72.8% on Orphan-335.
Interpretation
Enzyme function annotation and candidate enzyme screening can now be ordered by evidence chains rather than only by sequence or structural similarity; RIVER ranks possibly catalyzed reactions for a given enzyme and candidate enzymes for a given reaction. Earlier approaches relied on sequence or structural similarity or on training with known enzyme–reaction pairs, making it hard to integrate dispersed biochemical evidence across large candidate spaces and to handle enzyme families whose members are highly similar. In the E. coli in vitro metabolomics study reported by Gruber et al., 667 purified enzymes covered four experimentally supported enzyme–reaction associations, giving 2,668 candidate pairs; with Qwen3.6-35B-A3B, RIVER placed CobC, NanK and RihB first and DeoA third for a mean reciprocal rank of 0.833, and with GPT-5.6 Luna the four targets ranked 1, 1, 2 and 2 for a mean reciprocal rank of 0.750.
The ordering comes from staged evidence collection and review, and each candidate carries supporting rationales and unresolved questions. General-purpose coding agents can plan and call tools, but those abilities do not necessarily translate into better enzyme–reaction matching on this retrieval task. Under the same GPT-5.6 Luna foundation model, RIVER's mean reciprocal rank exceeded Luna plus Codex by 0.075 and the mean of repeated stand-alone Luna API runs (0.482) by 0.268; Claude Haiku 5.5 plus Claude Code reached 0.653, Phylo Max 0.708, Phylo Standard 0.310 and Enz-GFM 0.027.
The workflow first prescreens with evidence guidance, then separate agents interpret the reaction, organize protein annotations, check contradictions and review evidence before reranking candidates. Prescreening shrinks the candidate pool, so it must be checked for discarding correct candidates early. On P450, prescreening retained 51.1% of all candidate pairs while preserving 97.2% of labeled positive pairs, and 96.9% of queries kept at least one positive; on Phosphatase it retained 72.3% of candidate pairs and 79.6% of positive pairs, with every query keeping at least one labeled positive.
Perspective
The evaluation covers four experimentally supported cases from E. coli literature, five public benchmarks provided by EnzymeCAGE (Enzyme-405, Orphan-335, P450, Terpene and Phosphatase), and six cases of improved ranking across reaction types; a rank expresses relative priority within the supplied candidate pool, not a calibrated probability of catalytic activity.
RIVER's rankings do not establish catalytic activity or physiological relevance, and candidates lacking positive benchmark labels are not assumed to be experimentally inactive; prescreening retention varies considerably across families, the three families differ markedly in top-1% success rates and relative gains, and the relationship between the number of evidence categories and retrieval success is not monotonic for terpene synthases, while between-group differences are smaller for phosphatases.