{"paper_id":"dfdfbfc6-1977-4df3-95cc-750416b2f880","body_text":"Finding interesting phenomena in data is the essence of discovery. Yet interestingness remains a surprisingly elusive concept, requiring subjective human judgment and lacking the well-accepted metrics that concepts such as “statistical significance” enjoy.\nWe build a pipeline that extracts interesting hypotheses about connections between features and target diseases, including the direction of effect and potential underlying mechanisms. We identify three core concepts that lie at the heart of interestingness: novelty, utility (usefulness), and plausibility (the existence of an underlying explanatory mechanism).\nThe exponential growth of data and literature has not been accompanied by a corresponding growth in insights, and finding interesting, actionable insights from data remains a challenging task. Many now-obvious discoveries, such as the link between contaminated water and disease or handwashing, were overlooked for millennia. Hand hygiene gained acceptance only after germ theory, and H. pylori as the cause of ulcers was ridiculed until 1 ’s self-experimentation. Lithium, now essential for treating bipolar disorder would sound absurd if proposed naively 2 . These insights existed in the data but were missed or dismissed due to innate biases, insufficient explanatory frameworks or statistical rigor.\nThis work presents an integrative framework for quantifying and automating the discovery of interesting features in scientific datasets. We focus on identifying disease risk factors from the biomedical UK BioBank (UKB), although the underlying principles and methodologies are generalizable to other populations and non-medical domains. Our contributions are: We combine machine learning and natural language processing to create an expressive and easy to use pipeline (“InterFeat”) for finding interesting features. The InterFeat pipeline leverages structured data from electronic health records, biomedical ontologies and Knowledge Graphs ( KG ), scientific literature and large language models ( LLM s) to systematically identify, rank and explain features with high potential for discovery. We ground our approach in a formal definition of interestingness, integrating statistical and literature-based discovery approaches with LLMs to flexibly assess and rank features based on novelty, plausibility, and utility criteria, in relation to a target. We leverage LLMs to generate explanatory rationales and mechanistic explanations for candidate hypotheses , guiding researchers’ prioritization. We release our code and a novel expert-labeled multi-disease dataset of interesting biomedical features, with explanations and human validation. Using the UK Biobank (370,000+ patients), we demonstrate InterFeat’s effectiveness in uncovering previously undocumented risk factors across 8 major diseases, highlighting its potential to accelerate scientific discovery In temporal validation using a 2011 cutoff, over 8 diseases and 11,200 features, up to 21% of features retained after pipeline utility filtering were not reported in the literature until after 2011, demonstrating InterFeat’s ability to surface novel insights years before their documentation InterFeat surfaced features experts rated as interesting: across 137 candidates and 4 diverse, major diseases, 28% were judged interesting by physicians, and among top-ranked hypotheses InterFeat achieved 40-53% vs. 0-7% for a SHAP-only baseline. Figure 1 Left: InterFeat pipeline. (i) Target and features are extracted from a dataset (Here, the UK Biobank). (ii) Statistical and model-based methods are combined to retain features with predictive value (utility). (iii) UMLS-CUI linked entities are extracted and linked to a knowledge graph, to exclude known associations. Literature mining, via PubMed, removes frequent co-associations. Finally, (iv) language models (optionally augmented with relevant, retrieved texts) annotate the remaining features for novelty, plausibility and overall interestingness. Outputs include a ranked list of features with annotations and natural language explanations. Right: LLM Annotation Example. The input consists of a candidate feature-disease association. The LLM provides separate judgements (combined here for clarity) for novelty, plausibility, and overall interest, scored 1-4. This specific feature was confirmed as interesting, novel and useful by experts. Text edited for clarity.\nWe combine machine learning and natural language processing to create an expressive and easy to use pipeline (“InterFeat”) for finding interesting features.\nThe InterFeat pipeline leverages structured data from electronic health records, biomedical ontologies and Knowledge Graphs ( KG ), scientific literature and large language models ( LLM s) to systematically identify, rank and explain features with high potential for discovery.\nWe ground our approach in a formal definition of interestingness, integrating statistical and literature-based discovery approaches with LLMs to flexibly assess and rank features based on novelty, plausibility, and utility criteria, in relation to a target.\nWe leverage LLMs to generate explanatory rationales and mechanistic explanations for candidate hypotheses , guiding researchers’ prioritization.\nWe release our code and a novel expert-labeled multi-disease dataset of interesting biomedical features, with explanations and human validation.\nUsing the UK Biobank (370,000+ patients), we demonstrate InterFeat’s effectiveness in uncovering previously undocumented risk factors across 8 major diseases, highlighting its potential to accelerate scientific discovery\nIn temporal validation using a 2011 cutoff, over 8 diseases and 11,200 features, up to 21% of features retained after pipeline utility filtering were not reported in the literature until after 2011, demonstrating InterFeat’s ability to surface novel insights years before their documentation\nInterFeat surfaced features experts rated as interesting: across 137 candidates and 4 diverse, major diseases, 28% were judged interesting by physicians, and among top-ranked hypotheses InterFeat achieved 40-53% vs. 0-7% for a SHAP-only baseline.\nLeft: InterFeat pipeline. (i) Target and features are extracted from a dataset (Here, the UK Biobank). (ii) Statistical and model-based methods are combined to retain features with predictive value (utility). (iii) UMLS-CUI linked entities are extracted and linked to a knowledge graph, to exclude known associations. Literature mining, via PubMed, removes frequent co-associations. Finally, (iv) language models (optionally augmented with relevant, retrieved texts) annotate the remaining features for novelty, plausibility and overall interestingness. Outputs include a ranked list of features with annotations and natural language explanations. Right: LLM Annotation Example. The input consists of a candidate feature-disease association. The LLM provides separate judgements (combined here for clarity) for novelty, plausibility, and overall interest, scored 1-4. This specific feature was confirmed as interesting, novel and useful by experts. Text edited for clarity.\n\nAutomated Hypothesis generation  aims to systematize the traditionally intuitive process of discovery 3 , 4 . Methodologies, such as literature-based discovery (LBD), aim to identify missed connections between concepts and findings, thereby uncovering novel hypotheses 5 – 7 . However, traditional hypothesis generation approaches face several limitations: i) Directionality ignorance: These methods often treat associations between concepts as bidirectional, ignoring the direction of effect. For example, smoking reducing the risk of a disease would be novel, useful and interesting, while the inverse would not. ii) Ontology dependency: Many approaches rely on a standardized ontology and linkage to define co-occurrence, which limits it in terms of the source ontology and precision of linkage 7 . Recent studies have begun to address these limitations using deep learning and graph methods to improve the flexibility of LBD approaches 8 , 9 . Recent systems such as SciMON generate natural-language research directions based on retrieved literature, while knowledge-driven frameworks have advanced in biomedical tasks (e.g., heterogeneous-learning for drug repurposing, or research impact) 10 – 12 . InterFeat is complementary: it ranks existing hypotheses over observed features rather than free–form generation. Deep learning-based large language models have been used to automatically generate ideas and hypotheses and can flexibly capture unstructured relationships 3 , 4 , 8 , 13 – 15 . They have been shown to have near expert level scientific and medical understanding in some tasks, albeit when relying on known knowledge (e.g. differential diagnosis) 15 – 20 . However, their tendency to hallucinate unfeasible, or nonsensical ideas makes them insufficiently reliable to be a “Great Automatic Grammatizator” 21  for ideas without manual validation 22 . The use of actual features as a “starting point” may reduce hallucinations, due to the more limited hypothesis space 23 , 24 .\nIn practice, the starting point of many researchers looking for interesting connections in their data is statistical machine learning methodologies such as  feature selection 25 – 28 . Feature selection approaches focus on predictive power or statistical significance 29 – 31 . This includes the life sciences, such as predicting mortality 32 , Endometriosis 33 , protein function 34 , publication trends 11 , Heart attacks 35  and viral immune evasion 36 , 37 . There are many works using machine learning, and SHAPley values 38  have been applied to the UK Biobank to find risk factors, but most approaches rely on manual analysis of candidates, typically from a list of features sorted by model importance 28 , 33 , 39 – 41 . LLM-Select 25  used LLMs to select features by description and task, but again, only for predictive power.\n\nGiven a set of datasets over the same set of biomedical features and a target feature  y , our goal is to output a ranked list of interesting simple hypotheses of the form “ x  is related to  y , with a negative/positive correlation”, together with potential mechanisms underlying the hypothesis. For example, in the case of medical data, the target  y  often represents whether a patient will develop a specific disease. Features  x  are structured patient-level variables from the data, such as age, biomarkers (Vitamin D levels), genetic risk scores, questionnaires (smoking), medical history (age of asthma diagnosis, medications), etc.\nTo formulate a notion of interestingness, we are inspired by creativity literature, which frequently conceptualizes innovation as a confluence of  novelty  and  utility 42 – 44 . In other words, a creation is deemed innovative if it is both original and valuable or useful. Similarly, we define “interesting” hypotheses in scientific data if they satisfy the following criteria:  Novelty: \n x  should not be  established  in the literature or canonical knowledge bases as linked to the target  y . Alternatively, a hypothesis might be considered novel if there is a known connection between  x  and  y  but the direction of effect implied by the hypothesis is controversial or unestablished. A feature which is closely related to another known association may not be considered novel (e.g., cigarette vs. cigar smoke). Utility: \n x  must have predictive power, adding useful information for predicting  y . Note that in some use cases “utility” also implies that  x  needs to be actionable (e.g., smoking affects the risks of many diseases, and it can be changed). Plausibility:  In addition to the criteria inspired by creativity research, in science we believe another critical criterion is plausibility –  x  is consistent with current knowledge, and has a theoretical explanation. Medical data is particularly rife with spurious correlations, many of which are spurious or reflect underlying confounding factors. Thus, researchers tend to prioritize investigating correlations with plausible mechanisms. We operationalize these requirements by formalizing notions of novelty and utility, integrating well-known metrics (e.g., mutual information) with additional LLM input. LLMs also suggest mechanisms and explanations for each score. While we are not the first to combine notions of novelty and utility, we propose an integrative, configurable approach that we hope will be adapted by practitioners and serve as a vehicle for new scientific discoveries.\nNovelty: \n x  should not be  established  in the literature or canonical knowledge bases as linked to the target  y . Alternatively, a hypothesis might be considered novel if there is a known connection between  x  and  y  but the direction of effect implied by the hypothesis is controversial or unestablished. A feature which is closely related to another known association may not be considered novel (e.g., cigarette vs. cigar smoke).\nUtility: \n x  must have predictive power, adding useful information for predicting  y . Note that in some use cases “utility” also implies that  x  needs to be actionable (e.g., smoking affects the risks of many diseases, and it can be changed).\nPlausibility:  In addition to the criteria inspired by creativity research, in science we believe another critical criterion is plausibility –  x  is consistent with current knowledge, and has a theoretical explanation. Medical data is particularly rife with spurious correlations, many of which are spurious or reflect underlying confounding factors. Thus, researchers tend to prioritize investigating correlations with plausible mechanisms.\n\nOur pipeline is summarized in Fig.  1 . Here, we provide implementation details. Code and annotated datasets are available:  https://github.com/LinialLab/InterFeat . UKBB or SemMed raw data are unavailable due to licensing.\nImportantly, there are various ways to formulate novelty and utility. Our pipeline brings together the most prominent formulations, providing an intuitive way to configure and select those best suited for specific use cases.\nWe use the UK Biobank health records dataset as our main structured data source 45 . The dataset contains  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} 1681 patient covariates (medical record history, diagnostic results, medications, socioeconomic variables, genomic factors, lifestyle, etc.) measured at the time of each patient’s initial intake (2009 – 2011), with ICD-10 medical diagnoses recorded through 2022, for 370K adult patients. ICD-10/ICD-10-CM codes were also mapped to their phenotypes/Phecodes as additional covariates. For each target, we define the binary label y=1 if the target ICD code is first recorded after the participant’s intake; participants with the target diagnosis recorded on/before baseline are excluded, as are participants without any recorded diagnoses. We provide cross validated predictive results on different diseases, for the full cohort, IPW resampled data and ablation analysis of dropping InterFeat candidates ( Appendix A.6 ). This setup is used as a utility proxy for feature discovery as our focus is not a calibrated clinical model.\nWe clean and encode the raw UKB data into a structured format with  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} 3721 features. Features with missing values were mean-imputed, and a “missing” feature flag was added. Features without at least 30 non-missing values are dropped. Optionally, our pipeline removes redundant features using correlation feature selection. In interpretability use-cases, it is common to remove highly correlated features to reduce redundancy. A popular default is 0.8-0.95 for the Pearson correlation coefficient 27 , 46 . We use a 0.9 threshold, so that features with strong linear relationships are dropped as redundant, using the feature-engine library 47 .\nThe pipeline predicts whether a patient will be diagnosed in the future with a given disease (specified by ICD-10 medical codes). To help mitigate confounding by age, sex, and BMI, we optionally apply Inverse Propensity Weighting (IPW) on the negative samples 48 . The predicted probabilities are used as sampling weights for IPW, and the negatives are resampled down to a given ratio (9:1) ( 1 ).\nWe allow users to choose between several utility filters to remove features with no predictive strength for  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$y$$\\end{document} , each with a corresponding threshold. Specifically, p -value under a univariate test:  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{pVal}(x,y) \\;\\le \\; \\theta _{p}$$\\end{document} Mutual information between  x  and  y :  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{MI}(f,y) \\ge \\theta _{\\textrm{MI}}$$\\end{document} Model-based feature-importance score (e.g., global SHAP):  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{FImp}(f,y) \\ge \\theta _{\\textrm{FImp}}$$\\end{document} MI and FImp can ascertain non-linear effects. FImp reflects whether a feature is used by a trained predictive model(s), e.g., a boosting tree, unlike p-value. Users can choose criteria, thresholds and also whether to treat them as a conjunction (all) or disjunction (any). After some exploration, we chose lenient thresholds for our experiments:  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p\\text {-value} < 0.2$$\\end{document} ,  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\text {MI} \\ge 10^{-3}$$\\end{document} , or  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{FImp} \\ge 10^{-4}$$\\end{document} . In our selected configuration, a feature  x  passes the utility filter if it met any of the three criteria.\np -value under a univariate test:  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{pVal}(x,y) \\;\\le \\; \\theta _{p}$$\\end{document}\nMutual information between  x  and  y :  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{MI}(f,y) \\ge \\theta _{\\textrm{MI}}$$\\end{document}\nModel-based feature-importance score (e.g., global SHAP):  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\textrm{FImp}(f,y) \\ge \\theta _{\\textrm{FImp}}$$\\end{document}\nOur pipeline supports two ways to filter for novelty, both based on scientific literature.\nKG-based Filter.  We link features and target diseases to UMLS Concept Unique Identifiers (CUIs) 49  using scispaCy 50  and edges in SemMedDB v43 51 , a KG of 130 million semantic predications (subject–predicate–object triples) from 37 million PubMed citations. Features and targets are represented by sets of linked entities ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$E(x)$$\\end{document}  and  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$E(y)$$\\end{document} ), extracted using named entity recognition and linkage to entities in the KG (here, UMLS CUIs). If a feature  x  is already directly connected to the disease  y  in the KG (with sufficient evidence), we mark it as “known” (i.e.,  not novel ) and exclude it. To reduce the chance of false predicates, we filtered SemMedDB for predicates that had at least 2 unique citations as evidence, leaving 12.9 million. We treat the graph as unidirectional and ignore the type of predicate. scispaCy’s (V5.5) “en_core_sci_lg” entity recognition model was used, with a 0.88 threshold and 3 max entities per candidate, following recommendations for high-precision biomedical entity linking 50 , 52 . A predefined list of irrelevant high level categories are excluded by regex (e.g. “Qualification”, “Disease”, “Unit”).\nDomain-specific semantic similarity is computed between each feature and candidate entities, using a pretrained biomedical sentence-level language model (Biolord 53 ), measured as cosine similarity. This is used to further remove candidate entities with very low (defined as 0.1<) similarity to the feature, and later to define ‘strongly linked’ entities (e.g., ‘alcohol’ and ‘alcoholism’). Features were filtered out if all their linked entities were directly linked (1-hop) to the target(s) in the KG, or if they had at least one strongly linked entity ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\ge \\theta _{\\textit{sim}}$$\\end{document}  similarity) with a direct connection to the target(s). In our experiments, we chose a threshold of  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\theta _{\\textit{sim}}=0.4$$\\end{document}  as “strongly linked”.\nLiterature-based Filter:  Text mining is used to ascertain if the co-occurrence of features and disease is already established in the literature. This reflects the typical human search process: “Are there already papers about  x  and  y ?”\nPubMed is a large literature database of over 37 million published scientific and specifically biomedical works. We query the PubMed search API (including automatic term expansions) for publication counts of each feature, the target, and their co-occurrence ( x  AND  y ). If the pair is co-mentioned less than an absolute threshold  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\theta _{\\text {lit}}$$\\end{document}  or less frequently than expected by random chance (via one-way Fisher’s Exact Test,  p \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$<\\theta _{\\text {pval}}$$\\end{document} ), it is retained. Features with less than 20 hits in the database are left unfiltered (these could include, for example, recently coined terms). Again, after experimentation, we chose to use relaxed default threshold:  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\theta _{\\text {lit}} = 4$$\\end{document} ,  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\theta _{\\text {pval}} = 0.4$$\\end{document} .\nA note on thresholds.   In both novelty filters, we prefer high recall, filtering out only clearly non-novel features while retaining borderline cases. This prioritizes precision in exclusion, minimizing the risk of discarding under-explored but potentially meaningful findings. Consistent with this design, the initial utility screen is intentionally liberal to avoid premature exclusion; a post-hoc sensitivity analysis shows that the expert-validated discoveries are largely robust to substantially stricter statistical thresholds ( Appendix A.7 ).\nTo refine and rank filtered features, we use LLMs as an extra layer of information. Due to the nature of language models, we chose to focus on novelty and plausibility: language models are very effective for processing and internalizing vast amounts of (unstructured) existing knowledge, synthesizing multiple sources, and thus can often detect whether a certain hypothesis is already known. Similarly, their ability to integrate diverse pieces of knowledge and combine them in new ways helps them identify plausible mechanisms. We did not use the LLM to annotate utility, as this is something that is often use-case specific. Note that the LLM only annotates and explains candidate feature–disease pairs from a dataset (after any utility and novelty filters); it does not generate new candidates.\nWe annotate feature novelty and plausibility using GPT-4o-mini, selected after development-phase testing with Ai2’s OpenScholar, a LLaMA-3.1-8B variant 54 , 55 . This was motivated by GPT’s adherence to structured outputs. Chain of thought (COT) is used in all models’ prompts 56 . We use retrieval-augmented generation (RAG) 24 , using MedRag 57 , a biomedical retrieval toolkit, to retrieve related texts from the MedCorp corpus of 23 million PubMed abstracts, clinical textbooks and Wikipedia. The top 32 texts per feature and target, ranked by BM-25 are appended to the prompt. This outputs scored annotations and explanations. Local computation and data processing runs on consumer hardware in minutes to hours with 32GB RAM. the LLM annotation of thousands of candidate feature-–disease pairs in this study came to for under $50 total API spend. Each feature, target, their correlation and previous models’ explanations are run through the LLM (GPT-4o), to get an overall “Interestingness” (1-4) score and explanation. Prompts are provided in the appendix ( Appendix A.2 ), and outputs in the codebase. Finally, outputs are provided in a structured format for review, including annotation labels, Interestingness confidence score, feature statistics, and an explanation, sorted by confidence and feature importance. Due to the high cost of expert annotation, we limited validation to the first  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} 30 features with Interestingness  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$>2$$\\end{document} , sorted by score, per disease. See LLM output example in Fig.  1  (right). In this example, low Vitamin D levels increasing the risk of Esophageal cancer was rated as novel and moderately interesting (3/4). A mechanism from other cancers is noted, as is the unusual effect direction in this case. We note this feature was confirmed as interesting, novel and useful by annotators\n\nThe initial set of  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} 3721 features is filtered by the  utility  criteria, then further by the novelty and LLM steps, leaving less than  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} \n \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$2\\%$$\\end{document}  (under 80) final candidates per disease. This is consistent with other works suggesting examining up to 3% of features for hypothesis exploration in large, high dimensional data, notably the UKB 28 , 58 , 59 . Several observations from Table  1  are to be acknowledged: (i) The diseases span a wide prevalence range. The list includes rare diseases such as retinal vein occlusion (0.32%) but also high-prevalence diseases, such as depression (6.68%). (ii) The diseases cover cases of defined underlying biochemical mechanisms (e.g., gout) but also conditions without mechanistic explanation like depression. (iii) Some are early onset, while others are considered aging diseases. For example, coeliac is a lifelong autoimmune disease commonly diagnosed in childhood, while gallstones are more common in adults and can be treated. We conclude that these diseases display a reliable representation of other human diseases and conditions.\nWe observed that the number of features retained after the utility filter correlates positively with disease prevalence. This can be attributed to the fact that larger datasets, with more cases of a target in addition to background (negative/“healthy”) cases, provide greater statistical sensitivity to detect features with even modest associations. This effect is consistent with the UKB collected covariates, although diverse, being gathered under the assumption of their potential relevance to human health and wellness. Another observation concerns the knowledge graph (KG). For example, 72% of the features remained after KG filtration in the case of retinal vein occlusion, but only 40% for depression. Presumably, the richness of the KG is associated with the “popularity” of specific diseases 60 .\nTable 1 Pipeline statistics: Features retained at each stage per disease. Number of features kept by stage Target Disease Prevalence Utility Knowledge Literature Selected  Disease  Counts  (%)  Filter  Graph  Search  by LLM Cholelithiasis (gallstones) 19658 5.07 1447 697 157 50 Gout 9159 2.36 1707 812 148 62 Coeliac disease 2653 0.68 903 487 134 63 Spine degeneration 24867 6.42 2430 1187 136 73 Esophageal cancer 1518 0.39 611 408 152 59 Heart attack 3638 0.94 1008 520 102 43 Retinal vein occlusion 1246 0.32 558 402 163 60 Depression 28880 6.68 2537 1036 77 26\nPipeline statistics: Features retained at each stage per disease.\nEvaluating candidate hypotheses is challenging due to the difficulty in determining the accuracy of the hypotheses, and the intrinsic lack of a definitive ground truth for novel candidates. We assess our utility filters using time-stamped validation, an accepted methodology in hypothesis generation, when a definitive ground truth is unavailable 8 , 20 , 60 – 63 . In a nutshell, the idea is to take a cut-off date (in our case, 2011 – when the UKB study intake took place), and run the pipeline as if that date represents the present moment. This temporal validation evaluates only the utility-filtered features; we do not apply the SemMedDB/KG, literature or LLM steps in this experiment. SemMedDB is used only as a ground truth for when associations appear. For each of the 8 diseases, we took all the features that passed our utility filter and then examined whether those features were added as disease-associated entries in SemMedDB  after  2011. Since SemMedDB grows over time, a link appearing only post-2011 suggests our pipeline identified it prior to its recognition in the literature. Table  2  shows, per disease, how many of these  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} 11,200 discovered features were added in subsequent KG expansions, indicating that the pipeline can surface validated insights ahead of time. In particular, up to 21% of utility-filtered features appear in literature only after 2011. We found this reality check encouraging, as our utility filters were shown to retain valid features. Table 2 Temporal validation of utility filters by target disease. Target disease Total KG-linked features KG features (1-hop from target) Post-cutoff features Gallstones (cholelithiasis) 801 202 33 (16% [12–22]) Gout 920 274 58 (21% [17–26]) Coeliac disease 582 215 20 (9% [6–14]) Spine degeneration 1130 318 63 (20% [16–25]) Esophageal cancer 445 91 19 (21% [14–30]) Heart attack 643 320 18 (6% [4–9]) Retinal vein occlusion 400 10 0 (0% [0–28]) Depression 1214 537 60 (11% [9–14]) Statistics are provided for each target’s dataset of utility-filtered features. (i) total number of features linked to the KG, (ii) the number of features are directly connected (1-hop) to the target in the KG, and (iii) the count and percentage of features first reported after the temporal cutoff. Brackets indicate 95% confidence intervals (Wilson score).\nTemporal validation of utility filters by target disease.\nStatistics are provided for each target’s dataset of utility-filtered features. (i) total number of features linked to the KG, (ii) the number of features are directly connected (1-hop) to the target in the KG, and (iii) the count and percentage of features first reported after the temporal cutoff. Brackets indicate 95% confidence intervals (Wilson score).\nOur primary question is whether the pipeline’s outputs are indeed interesting, according to domain experts. We performed a focused human evaluation on:  Gout ,  Cholelithiasis (Gallstones) ,  Esophageal cancer  and  heart attacks . We aimed to (i) measure alignment between expert and pipeline judgments, and (ii) assess whether experts found value in the pipeline’s discoveries.\nFour senior medical doctors, each with over 10 years research experience, including with these diseases, annotated 137\npipeline-selected features for  novelty ,  plausibility ,  utility , and overall  interestingness , on a 1-4 scale with explanations. The challenging and ambiguous nature of the task demanded domain knowledge, and each target was reviewed by an expert on it.\nOf the features marked as interesting by the models, up to 42 candidates per disease were selected by the confidence score, as given the constraints of manpower and costs, a full-scale evaluation was not feasible. For heart attacks, only the top model-candidates were annotated. Scores were binarized ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$>2$$\\end{document} ) when comparing with model annotations.  Overall, 28% of candidates were interesting to the doctors: 18% of Gout, 30% of Esophagus cancer and 37% of Cholelithiasis .\nOn binarized scores, the pipeline agreed with experts in 40% of cases for  novelty , 57% for  plausibility , 79% for  utility , and 69% for overall  interestingness . When evaluating the raw  1–4 scores  using Cohen’s Kappa,  plausibility  showed fair agreement ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\kappa = 0.29$$\\end{document} ). However,  utility  ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\kappa = -0.10$$\\end{document} ) and  interestingness  ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\kappa = 0.009$$\\end{document} ) showed low alignment, underscoring the inherent subjectivity of a continuous scoring criteria.\nTo evaluate the expert annotators’ ability to distinguish meaningful features from distractors, we added  distractor features  into each annotation dataset. These features were derived by randomly sampling from those discarded which did not pass the utility filter. This helped assess annotator bias and task difficulty. For each target we added 20% distractors, yielding 35 total additional annotation candidates, in addition to the original, real features. Annotators were not informed of the distractors. GPT-4o was prompted to generate justifications for why each distractor was interesting ( Appendix A.2 ). It has been shown that LLMs can fool humans in such scenarios 64 . Statistical comparisons were performed using two-sample t-tests, summarized in Table  3 . Human annotators recognized the distractors as having lower plausibility, utility and interestingness. Table 3 Comparison of human annotations between real and distractor (dist.) features. Annotation Mean (real) Mean (Dist.) p-Value Novel 2.78 2.82 0.83 Plausibility 2.46 2.12 0.04 Utility 1.94 1.48 0.0005 Interestingness 2.09 1.81 0.04\nComparison of human annotations between real and distractor (dist.) features.\nTo evaluate InterFeat’s ability to identify interesting features compared to a baseline of selecting by feature importance, we compared (and annotated) the top 15 candidate features generated by the pipeline as well as its individual components for Gallstones, Esophageal Cancer, and Gout. Table  4  summarizes the number of features validated as interesting for each approach, out of the top 15, sorted by SHAP. SHAP 38  is a popular method for identifying feature importance, and reflects a typical data scientist or computational researchers’ likely default.\nSHAP shows which features drive model predictions, including the direction of effect and in relation to other features’ contributions, in a consistent framework. SHAP based methods have been extensively applied, including on the UKB 32 , 33 , 35 , 41 , 65 , making it a natural comparison for getting a starting list of features to analyze, as in 28 .\nThe methods compared include: the SHAP baseline, representing feature selection based solely on predictive importance; an L1-regularized linear model baseline 32 , 66 ; intermediate filters (Knowledge Graph (KG) only, Literature only, and combined KG+Literature); the full InterFeat pipeline; and an additional experimental step (“InterFeat + ReasonLM”). This extra step reranked all InterFeat selections simultaneously using a separate, reasoning LLM (Google Gemini 2.5 Pro 67 ), allowing for list-wise reranking; it serves here primarily for analytical comparison and is not part of the standard pipeline. All candidates are still filtered for utility, then sorted for top 15 by feature importance. As shown in Table  4 , InterFeat consistently identified more interesting features than the feature importance baselines across all targets (e.g., InterFeat vs. SHAP/L1: 6 vs 1/3; 5 vs 0/1; 3 vs 0/2). This difference was statistically significant for the three diseases in aggregate against both the SHAP baseline (Fisher’s exact test, two-sided,  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p=0.0003$$\\end{document} ) and the L1 baseline ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p=0.043$$\\end{document} ,  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$n=90$$\\end{document} ). Annotations available in  Appendix A.4  and repository (“Ablation Results”). Table 4 Comparison of validated interesting features by method. Method Gallstones Esoph. Ca. Gout SHAP Baseline 1 0 0 L1 Baseline 3 1 2 KG 2 0 0 Literature 3 0 0 KG+Literature 5 1 3 InterFeat 6 5 3 InterFeat+ReasonLM 6 10 5 Results for top 15 (sorted by SHAP). All methods include utility filtering.\nComparison of validated interesting features by method.\nResults for top 15 (sorted by SHAP). All methods include utility filtering.\nRanking metrics . We computed ranking metrics on the real, annotated InterFeat candidates. Candidates were ranked by the LLM’s overall confidence score and compared to the experts’ 1–4 “Interestingness” ratings. We report NDCG and MRR. Results (Table  5 ) show high NDCG values across all targets , indicating that the LLM’s ordering aligns well with expert judgement. Table 5 Ranking quality evaluation on the InterFeat real candidate set. Target R NDCG MRR Cholelithiasis–Gallbladder 16 0.906 0.250 Gout 9 0.860 0.167 Oesophagus cancer 21 0.889 0.333 Only candidates with expert annotations are included ( R  counts unique features with human rating  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\ge 3$$\\end{document} ). NDCG uses graded relevance (expert ratings 1–4). MRR uses a binary relevance threshold (ratings  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\ge 3$$\\end{document} ).\nRanking quality evaluation on the InterFeat real candidate set.\nOnly candidates with expert annotations are included ( R  counts unique features with human rating  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\ge 3$$\\end{document} ). NDCG uses graded relevance (expert ratings 1–4). MRR uses a binary relevance threshold (ratings  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\ge 3$$\\end{document} ).\nSensitivity and Robustness . We conducted a retrospective sensitivity analysis to evaluate hyperparameter choices (Table S4). The utility criterion results demonstrate high robustness: while our default configuration uses a liberal statistical threshold ( \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p < 0.2$$\\end{document}  OR any other utility criteria (e.g. MI, SHAP)) to maximize recall, tightening this threshold to a stringent  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p < 0.01$$\\end{document}  still retained the majority of expert-validated interesting features (Recall: 65.4% vs 84.6% at  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p<0.2$$\\end{document} , and 53% at  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$p < 0.001$$\\end{document} ). This indicates that most novel associations identified by the pipeline are strong statistical signals rather than borderline cases dependent on a loose filter.\nOf 375 features marked as interesting by LLMs across all 8 targets, 48% were picked more than once, with 6 appearing in 6+ of the targets: ’melanoma genetic risk’, ’Microalbumin in urine’, intraocular pressure genetic risk’, ’Arm fat percentage’, ‘epithelial ovarian cancer genetic risk’, ’age at menopause genetic risk’. These may highlight underlying factors such as genetics or immunology that may affect many diseases 68 . Not all causes of diseases are understood, and some may have multiple etiologies 20 , 68 . Furthermore, variables such as age, obesity or inflammation can drive conditions without implying direct causal links, and may reflect more fundamental factors that predispose to diseases. For instance, high arm fat percentage relates to confounders such as muscle mass, BMI and general frailty. We acknowledge that these might be caused by confounders rather than truly novel or causal effectors, although this does not necessarily affect utility 69 . We grouped features into semantic categories, using a combination of manual annotation and LLM-assisted clustering (see Appendix, Fig.  2 ).\nInterFeat selected hypotheses validated as particularly interesting by annotators included:\nEsophageal cancer is an aggressive malignancy, defined by ICD-10 code C15. It has  \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\sim$$\\end{document} 81K PubMed publications but is relatively rare in the UKB due to low survival rates. Genetic Risks associated with other diseases:  melanoma, ischemic stroke, rheumatoid arthritis, systemic lupus erythematosus. The association with melanoma suggests shared genetic or inflammatory pathways. Genetic risks linked to rheumatoid arthritis and lupus indicate that autoimmune and inflammatory processes could play a role in esophageal cancer 70 . Asthma diagnosis and genetic risk:  Possibly linked via chronic inflammation or steroids 71 . Atenolol:  a beta-blocker for cardiovascular disease. Epithelial Ovarian Cancer genetic risk  exhibited a particularly interesting negative association. Novel Biomarkers:  Vitamin D, Acetoacetate, Acetone.\nGenetic Risks associated with other diseases:  melanoma, ischemic stroke, rheumatoid arthritis, systemic lupus erythematosus. The association with melanoma suggests shared genetic or inflammatory pathways. Genetic risks linked to rheumatoid arthritis and lupus indicate that autoimmune and inflammatory processes could play a role in esophageal cancer 70 .\nAsthma diagnosis and genetic risk:  Possibly linked via chronic inflammation or steroids 71 .\nAtenolol:  a beta-blocker for cardiovascular disease.\nEpithelial Ovarian Cancer genetic risk  exhibited a particularly interesting negative association.\nNovel Biomarkers:  Vitamin D, Acetoacetate, Acetone.\nGallstones, or cholelithiasis, are a prevalent hepatobiliary disorder, with 101K publications, characterized by the formation of calculi within the gallbladder, defined by the ICD-10 range K80-K82. Pharmacological Influences:  Omeprazole, a proton pump inhibitor. These drugs have been claimed to affect gallbladder function 72 Genetic Risks associated with other diseases:  such as breast cancer, primary open-angle glaucoma, Alzheimer’s, and schizophrenia. The association between  breast cancer genetic risk  and gallstones may reflect shared metabolic pathways. Lipid Metabolism Markers:  Apolipoprotein B/A1 ratio, Medium HDL cholesterol. The ApoB/ApoA1 ratio, indicative of lipid metabolism balance, reinforces the role of lipid dysregulation in gallstone pathogenesis. These suggest therapeutic strategies aimed at regulating lipid profiles. Psychiatric Conditions:  Bipolar disorder, depression, neuroticism. May indicate a systemic metabolic factor or medication effect.\nPharmacological Influences:  Omeprazole, a proton pump inhibitor. These drugs have been claimed to affect gallbladder function 72\nGenetic Risks associated with other diseases:  such as breast cancer, primary open-angle glaucoma, Alzheimer’s, and schizophrenia. The association between  breast cancer genetic risk  and gallstones may reflect shared metabolic pathways.\nLipid Metabolism Markers:  Apolipoprotein B/A1 ratio, Medium HDL cholesterol. The ApoB/ApoA1 ratio, indicative of lipid metabolism balance, reinforces the role of lipid dysregulation in gallstone pathogenesis. These suggest therapeutic strategies aimed at regulating lipid profiles.\nPsychiatric Conditions:  Bipolar disorder, depression, neuroticism. May indicate a systemic metabolic factor or medication effect.\nBiomarkers:  Direct bilirubin, Acetoacetate and Acetone. Long-term or frequent  childhood antibiotic use . May relate to microbiome. Increasingly validated by recent works 73 , 74 Higher lean leg mass: counterintuitively associated with increased risk. Anxiety or panic attacks linked to higher risk. Post-traumatic stress disorder to  lower  risk.\nBiomarkers:  Direct bilirubin, Acetoacetate and Acetone.\nLong-term or frequent  childhood antibiotic use . May relate to microbiome. Increasingly validated by recent works 73 , 74\nHigher lean leg mass: counterintuitively associated with increased risk.\nAnxiety or panic attacks linked to higher risk. Post-traumatic stress disorder to  lower  risk.\n\nWe present an integrative pipeline that combines statistical feature selection, knowledge-graph screening, and retrieval-augmented LLM annotation to discover interesting features, defined as a combination of  novel, plausible, and having utility . Our approach systematically narrows thousands of raw features to a concise shortlist. Compared to ranking features solely by statistical or model importance measures (e.g., SHAP values), we demonstrate superior performance. 40–-53% of the top 15 candidates per target were validated as interesting, compared to a 0-–20% rate for the SHAP or L1 baselines’. For instance, InterFeat identified ’long-term childhood antibiotic use’ as a risk factor for myocardial infarction. Recent works have since supported this association 74 , 75 . This supports the pipeline’s capacity to find signals before the clinical literature.\nDespite progress, challenges remain. Imperfect knowledge bases’ coverage can lead to features being falsely labeled as novel, and LLM judgment may still misalign with humans. Future work could explore ablations of pipeline components, incorporate additional criteria for interestingness to improve alignment with human judgment, and develop more sophisticated ways of fusing structured meta-data with the LLM. Improved integration of feature attributes may also help identify novelties based on unusual population subsets or non-monotonic effects, moving beyond the existing usage of directionality. Future iterations could explicitly model feature interactions 32 , 33 , 66 , 76 . Ongoing improvements in large language models suggests that some filtering stages–such as the knowledge-graph pass–could be removed at the cost of higher LLM compute costs 77 , 78 . Exploring this trade-off is outside our present scope but forms a natural direction for follow-up work. We plan to apply our pipeline on a large scale to hundreds of major diseases, providing the candidates as a community resource. Although the alignment of pipeline scores with human assessments for the top-ranked subset of candidates is modest, it is crucial to note this subset is distilled from an initial pool of thousands. Generating a ranked list of candidates enriched for interestingness improves on standard practices (e.g., ranking by predictive importance, or manual review of every single hypothesis), offering clear value as a time-saving tool for researchers and a starting point for expert validation. Our approach is flexible, and outputs a ranked, grounded set of interesting features at scale, with higher enrichment of validated “interesting” hypotheses than the utility baselines in our evaluation, while avoiding the risk of ungrounded generation (“hallucinations”) by avoiding free-form generation entirely. Our approach is generalizable to other domains, and we look forward to expanding it, improving AI-human alignment in formulating what is interesting.","source_license":"CC-BY-4.0","license_restricted":false}