Their system, called React-to-Me, is a conversational assistant built on top of the Reactome Pathway Knowledgebase, one of the most rigorously maintained curated resources in molecular biology. Reactome, stewarded by a team of expert biologists, documents human biological pathways, molecular interactions, and disease mechanisms with a level of care that ordinary web search cannot match. Yet the knowledgebase has a usability problem: its data model is intricate, and its search interface rewards people who already know the vocabulary and structure of the underlying database. For a graduate student, a clinician, or a researcher crossing into a new field, that learning curve can be discouraging. React-to-Me was designed to let such users ask questions in plain natural language while still receiving answers that are traceable, verifiable, and anchored in curated content.

The engineering heart of the system is a hybrid retrieval-augmented generation pipeline, a technique that constrains a language model to answer only from documents fetched at query time rather than from its internal memory. In React-to-Me, an incoming question triggers two complementary searches over the Reactome corpus. The first is a semantic vector search, which uses neural embeddings to match the meaning of the question against the meaning of database entries, capturing paraphrases and conceptual overlap that keyword matching would miss. The second is a lexical keyword search, of the BM25 variety familiar from classic information retrieval, which excels at pinning down precise identifiers, gene names, and chemical terms. The two ranked lists are merged with reciprocal rank fusion, a method that rewards documents appearing near the top of both lists, and the fused evidence is handed to a language model that is instructed to generate an answer grounded strictly in that retrieved material.

Grounding alone is not enough; the system also insists on provenance. Every response is directly linked to the corresponding Reactome entries from which it was assembled, so users can click through and inspect the curated pathways behind the answer. And when the retrieved evidence is thin or absent, React-to-Me is designed to defer rather than improvise: it declines to speculate and instead points users toward trusted external biomedical sources. This refusal behavior, the authors argue, is a feature rather than a failure, because an honest admission of limited coverage is far more valuable to a scientist than a fabricated pathway. The team tested how well that refusal logic worked in the wild by manually classifying logged instances where the system declined to answer, finding that out-of-scope biology questions were the largest category, accounting for 31.9 percent of refusals.

Before any human saw the system, the researchers put the retrieval engine itself through computational benchmarking using the Ragas evaluation framework, scoring configurations on context utilization, relevance, and faithfulness across several query types: simple factual lookup, mechanistic reasoning, multi-context synthesis, and conditional logic. The results showed a clear pattern. Semantic retrieval on its own preserved relevance but faltered on faithfulness, letting unsupported statements slip through, while pure keyword retrieval weakened on complex reasoning tasks that required synthesizing multiple sources. The combined hybrid approach, ranked by reciprocal rank fusion, delivered the highest and most stable scores across every metric and every query type, confirming that the two retrieval strategies compensate for each other’s blind spots.

Real-world deployment brought a second wave of evidence. Analyzing user queries logged during October 2025 and March 2026, the team found that factual lookup and mechanistic or relational biology questions dominated, together accounting for 56.1 percent of classified logged-in queries. Notably, the mix shifted over time: factual lookups rose from 24.7 percent to 33.7 percent of classified queries, while mechanistic and relational questions declined from 32.7 percent to 21.2 percent, a pattern suggesting that as users became comfortable with the tool, they leaned on it increasingly as a rapid reference rather than a reasoning partner. The study received ethics approval from the University of Toronto Research Ethics Board under protocol number 47192, and all survey participants provided informed consent.

The most demanding test was a blinded head-to-head comparison against a general-purpose model, GPT-4o-mini, judged by ten external molecular biology experts who did not know which system produced which answer. The verdict favored grounding decisively. Grounded React-to-Me responses were more likely to receive higher quality ratings than their ungrounded counterparts, with an overall odds ratio of 2.01. The gains were strongest where scientific rigor matters most: factual accuracy improved with an odds ratio of 2.97, biological specificity with an odds ratio of 2.88, and mechanistic depth with an odds ratio of 1.83. Individual evaluator preferences, recorded in the supplementary data, showed consistent majority favoring of the grounded system across question-level comparisons, an exploratory mixed-effects ordinal regression accounting for variation among evaluators and questions.

User sentiment reinforced the expert verdict. In a survey of 25 consenting users drawn from the publicly deployed system, 92 percent expressed strong satisfaction with ease of use, 88 percent with citation reliability, and 85 percent with factual accuracy. Strikingly, perceived accuracy showed the strongest association with overall confidence in the system, with a correlation of r = 0.81, meaning that users who trusted the answers did so primarily because the facts checked out, not because the interface was pretty. Respondents spanned organization types and research fields, with human biology research most represented, followed by computational biology, cell biology, and genetics, and satisfaction held up across education levels and levels of prior Reactome experience.

Beyond the headline numbers, the study offers a practical template for institutions wrestling with how to deploy AI responsibly. The team published monthly operating costs, the full survey instrument, benchmark questions, and analysis code as supplementary materials, an unusual degree of transparency for a production AI deployment. The willingness to defer to trusted sources when coverage runs out, the insistence on clickable citations for every claim, and the hybrid retrieval architecture that balances conceptual breadth with lexical precision together sketch what credible scientific AI might look like. As funders and publishers grapple with hallucination risks in research tooling, React-to-Me demonstrates that domain-specific grounding is not a theoretical nicety but a measurable design choice, one that nearly triples the odds that an expert will judge an answer factually accurate. The broader lesson may extend well beyond pathways: for specialized knowledge, the future of conversational AI belongs to systems that know the boundaries of what they can honestly say.

Subject of Research: A grounded conversational AI interface for querying the Reactome Pathway Knowledgebase

Article Title: React-to-Me: real-world experience of a grounded conversational interface to the Reactome Pathway Knowledgebase

Article References: Mohammadi, H., Almodaresi, F., Hogue, G. F. J., Wright, A., Orlic-Milacic, M., Li, N. T., Mawani, A., & Stein, L. (2026). React-to-Me: real-world experience of a grounded conversational interface to the Reactome Pathway Knowledgebase. BMC Biology. https://doi.org/10.1186/s12915-026-02740-2

Image Credits: AI Generated

DOI: 10.1186/s12915-026-02740-2

Keywords: Reactome Pathway Knowledgebase, conversational AI, retrieval-augmented generation, large language models, knowledge grounding, biological pathways, biomedical question answering, hybrid retrieval, language model evaluation, computational biology, expert evaluation, responsible AI

Metabolomics, the large-scale study of small molecules such as lipids, amino acids and sugars that circulate in the blood, has become a cornerstone of modern epidemiology. Modern high-throughput nuclear magnetic resonance (NMR) platforms can quantify hundreds of metabolites in a single serum sample, and researchers typically interrogate these data one metabolite at a time, testing whether the average level of each molecule shifts in response to a disease, exposure or intervention. This univariate framework, exemplified by metabolome-wide association studies, has produced valuable biomarkers for conditions ranging from diabetes to cardiovascular disease. But it has a structural blind spot: it treats each metabolite as an independent entity, ignoring the dense web of correlations that connects them through shared biochemical pathways, shared genetics and shared environmental influences.

The Bristol team, led by Rachel M. Wood and Laura J. Corbin, with statistician Daniel J. Lawson as senior author, set out to address that blind spot directly. Their tool of choice, CLARITY, is a non-parametric method designed to compare two datasets describing the same entities and to quantify whether the similarities between those entities are conserved between the two. In its previous applications, CLARITY had been used mainly for exploratory anomaly detection on synthetic bioinformatics data. The new study marks its first serious deployment on real clinical metabolomics data, with the shared entities redefined as metabolites themselves rather than individuals.

The clinical setting was the By-Band-Sleeve trial, a pragmatic randomised controlled trial conducted in the UK that compared three forms of metabolic and bariatric surgery: Roux-en-Y gastric bypass, adjustable gastric banding and sleeve gastrectomy. The researchers focused on a pilot subset of 125 patients recruited at a single hospital site for whom serum samples were available both before surgery and at 36 months after randomisation. Those 250 samples were analysed on a high-throughput 1H-NMR platform operated by Nightingale Health in Helsinki, yielding 250 metabolic biomarkers, of which 160 were absolute concentration measures used in the final analysis. After rigorous quality control using the R package metaboprep, which removed samples and features with excessive missingness and outliers identified by principal component analysis, data from 237 samples and 121 individuals remained.

The analytical strategy proceeded in two parallel streams. First, the team fitted a linear mixed model to each metabolite, treating timepoint as the predictor and including storage time, age and sex as covariates, with participant identity as a random effect to account for the repeated measures on the same individuals. This traditional approach asks a simple question: did the average level of this metabolite change after surgery? Benjamini-Hochberg correction was applied to control the false discovery rate across the many simultaneous tests. Second, and more innovatively, the team ran the CLARITY pipeline, which works very differently. CLARITY builds a representation of the covariance structure among metabolites at baseline and asks how well increasingly complex models built from that baseline can predict the covariance structure at follow-up. Metabolites whose relationships with the rest of the metabolome cannot be predicted from baseline data, and for whom this unpredictability persists across a wide range of model complexities, are flagged as anomalies and assigned a persistence score.

The results from the two approaches were revealing in their partial overlap and partial divergence. The linear mixed model identified 30 metabolites, roughly 19 percent of those tested, as significantly altered by the intervention. The strongest signal came from triglycerides in large HDL particles, which increased by 0.78 standard deviations after surgery, alongside notable increases in HDL particle diameter and striking decreases in the branched-chain amino acids valine and isoleucine. That last finding aligns with a substantial body of literature linking bariatric surgery to reduced circulating branched-chain amino acids, changes thought to contribute to the metabolic benefits of the procedure.

The CLARITY analysis told a complementary story. Twenty-eight metabolites, about 18 percent, showed statistically meaningful persistence, meaning their relationships with other metabolites had been restructured in ways that baseline data could not predict. Dominating this list were components of very large HDL particles: particle concentration, total cholesterol in large HDL, cholesterol esters, total lipids and phospholipids. Crucially, only some of these overlapped with the metabolites flagged by the mixed model. APOA1, the principal protein component of HDL, showed clear covariance restructuring, as did several medium and large HDL species. Meanwhile, the branched-chain amino acids, so prominent in the univariate analysis, showed little persistence at all, indicating that their mean levels had changed through a pre-existing pathway without forming any new relationships with the wider metabolome.

To validate these interpretations, the team constructed a simulation framework mimicking the essential features of real NMR metabolomics data, including clusters of covarying metabolites, observed and unobserved confounders, and planted interventions that either strengthened an existing pathway or activated an entirely new one. In this simulated world where the ground truth is known, the two methods proved genuinely complementary. Metabolites experiencing a direct shift in mean levels were best detected by the linear mixed model, which predictably lost all power when the mean shift approached zero. CLARITY, by contrast, reliably detected metabolites whose covariance structure had changed, including a subtle but biologically important category: metabolites that experienced no direct change themselves but whose relationships shifted because their partners in a pathway had been altered. The team describes these as second-order changes, invisible to any regression on individual metabolites but clearly legible in the relational structure. Combining both methods produced the best classification of anomaly types across all simulation scenarios, with CLARITY adding value even when the univariate signal was strong.

The biological implications of the covariance findings are worth pausing on. The metabolites flagged most strongly by CLARITY were overwhelmingly HDL-related, and this resonates with a growing consensus in cardiovascular research that the quality and functionality of HDL particles, determined by their lipid and protein composition, shape and number, matter more to health than their sheer quantity. If bariatric surgery does not merely raise HDL levels but reorganises the relationships among HDL components and the wider lipidome, that structural change may encode information about genuine pathway activation that a simple concentration measurement cannot capture. The authors are careful to note that their results are exploratory and the contribution primarily methodological, but the pattern is nonetheless suggestive of a rewiring of lipid metabolism that standard analyses would never reveal.

The team also conducted a series of sanity checks to rule out mundane explanations for their findings. They computed the total correlation, the sum of absolute pairwise correlations, for each metabolite at both timepoints and confirmed that the persistent anomalies were not simply metabolites whose overall connectivity had uniformly increased or decreased. Nor was persistence driven by baseline connectedness or by unusually variable measurements. The cross-validation procedure used to derive empirical p-values appears to effectively account for such features, giving confidence that the anomalies genuinely reflect changes in metabolite covariance rather than technical artefacts.

The broader significance of this work extends well beyond bariatric surgery. Metabolomic data are notoriously redundant and noisy, with unobserved factors such as fasting status, exercise, sleep and time of day all leaving their fingerprints on measured levels, and these confounders vary unpredictably between sampling occasions. Methods that examine each metabolite in isolation conflate these technical and environmental shifts with genuine biological change. CLARITY sidesteps much of this difficulty because it is deliberately insensitive to uniform scaling of metabolite levels and to changes in large correlated baseline clusters, which would swamp univariate comparisons. It requires no training data, uses only covariances, and could in principle be applied even when the two datasets come from different individuals, opening the door to comparisons across cohorts, treatments or disease states.

The authors are candid about the method’s limitations. Because CLARITY corrects for structure learned from the baseline dataset, it has limited power to detect changes in factors already present at baseline, and there is as yet no theoretical guidance on optimal pre-processing choices, though the team’s sensitivity analyses suggest that scaling choices do not compromise validity. They point toward related techniques such as unfolded spectral embedding as complementary tools for quantifying changes in existing structures. The next steps they envision include comparing patient metabolomes against a healthy reference profile, contrasting different surgical procedures head to head, and pairing the statistical anomalies with laboratory experiments that can probe the actual pathways activated.

What the study ultimately demonstrates is a principle: the co-occurrence patterns of metabolites carry information about biological pathways that individual measurements do not, and tools exist to read that information. As multi-omic datasets grow ever larger and clinical interventions are increasingly evaluated through molecular lenses, methods like CLARITY may become essential for translating the flood of high-dimensional data into genuine mechanistic insight about how the body responds to treatment.