Ilias Tagkopoulos

Despite the growing scale of untargeted food metabolomics, scalable approaches for translating raw LC-MS profiles into functionally prioritized, experimentally testable candidates remain limited. Most detected features remain structurally unresolved, and the few that are annotated cannot be reliably connected to whole-food biological activity through additive models alone. Here we present a multi-layer prioritization framework that addresses this bottleneck across 322 commonly consumed foods in the United States, comprising 30,343 small molecules. A retention-index model ( R ² = 0.95, PCC = 0.97) expanded putative structural coverage from 7.4% to over 78% of detected features, assigning confidence-tiered structural proxies through formula-constrained matching against 11.8 million PubChem compounds. A two-stage large language model workflow harmonized 39,158 assay protocols spanning 119,633 chemical-assay measurements from ChEMBL into standardized bioactivity classes, enabling compound-level evidence to be propagated across the food composition space. The framework was applied to antioxidant activity as a primary, well-characterized endpoint while additional bioactivity domains (anti-inflammatory, anticancer, antidiabetic, antiviral, and antibacterial effects) were incorporated as compound-level evidence layers to support metabolite prioritization. Integrating concentration-aware activity mapping with a machine learning model trained on whole-food antioxidant measurements yielded strong predictive performance (R² = 0.73, PCC = 0.85) and prioritized metabolites most strongly associated with model-predicted food-level antioxidant capacity via SHAP attribution. Structure-based analysis using Boltz-2 against the antioxidant sensor provided a final mechanistic triage layer, distinguishing high-priority candidates from lower-ranked compounds. Applied across 322 foods, this funnel reduces tens of thousands of untargeted LC-MS features to a confidence-ranked shortlist for targeted wet-lab validation and provides a generalizable strategy for translating food metabolomics into biologically interpretable and actionable hypotheses. This framework establishes a scalable foundation for transforming food composition data into actionable biological insight, advancing the integration of food, health, and data science. More broadly, these results highlight limitations of traditional nutrient-centric frameworks and support a shift toward modeling molecular patterns and compositional signatures to better capture the functional properties of foods.

The rapid emergence of multidrug-resistant bacteria has created an urgent need for improved antimicrobial discovery and screening platforms. Here, we present ARCADIAMP, a generative and virtual screening platform that couples an iterative-learning discrete denoising diffusion probabilistic model with a two-stage Evolutionary Scale Modeling 2 (ESM2)-based antibacterial activity classifier to generate, classify, and prioritize potent AMPs with high activity, low toxicity, and favorable serum stability. Eight of the ten experimentally screened peptide candidates showed antimicrobial activity (MIC ≤ 32 μg/mL), while one generated candidate, Arcinin, demonstrated strong activity against ESKAPE pathogens (MIC 8–32 μg/mL), low hemolytic activity (LC ₅₀ > 512 μg/mL for human red blood cells), and strong serum-retained activity (MIC 32 μg/mL in 50% bovine serum for four ESKAPE species). Electron microscopy, membrane depolarization assays, time-kill kinetics, and molecular dynamics simulations showed that Arcinin acts through sub-microsecond insertion and penetration consistent with the behavior of other well-known AMPs. In a bacteria-infected wound murine model, Arcinin achieved a 4-log reduction in bacterial burden, which facilitated subsequent re-epithelialization and wound recovery. By framing antimicrobial discovery as an AI-assisted iterative optimization problem, ARCADIAMP links activity, toxicity, and efficacy and provides a scalable template for discovering therapeutically promising biologics.

Global food systems must deliver nutritious, sustainable foods while sharply reducing environmental impact. Yet, food innovation remains slow, empirical and fragmented. Artificial intelligence (AI) offers a transformative path to link molecular composition to functional performance, connect chemical structure to sensory outcomes and accelerate cross-disciplinary innovation across the production pipeline. While it is broadly applicable to food systems, we focus on sustainable proteins-plant-based, fermentation-derived and cultivated-as a high-impact test bed for AI-driven closed-loop design. We review the applications, opportunities and challenges of AI for food as an emerging discipline that integrates ingredient design, formulation development, fermentation and production, texture analysis, sensory science, manufacturing and recipe generation. We identify four priorities: advancing scientific machine learning with embedded domain priors, treating food as a programmable biomaterial, building self-driving laboratories for automated discovery and developing deep reasoning models that integrate nutrition and sustainability. Integrating AI responsibly into the food innovation cycle can accelerate the transition to sustainable food systems and establish a predictive, design-driven science of food for human and planetary health.

An important goal for personalized diet systems is to improve nutritional quality without compromising convenience or affordability. We present an end-to-end framework that converts dietary standards into complete meals with few ingredient changes. Using the What We Eat in America (WWEIA) intake data for 135,491 meals, we identify 34 interpretable meal archetypes that we then use to condition a generative model and a portion predictor to meet USDA nutritional targets. In comparisons within archetypes, generated meals show a median 47.0% reduction in absolute deviation from per-meal RDI targets, while remaining compositionally close to real meals. Our results show that by allowing one to three food substitutions, we were able to create meals that were 10% more nutritious, while reducing costs 19–32%, on average. By turning dietary guidelines into realistic, budget-aware meals and simple swaps, this framework can support public-health programs and consumer apps; clinical decision support is a promising future direction pending further validation and safety review, to deliver scalable, equitable improvements in everyday nutrition.

Understanding and predicting consumer acceptance is critical to commercial success in the coffee industry. This study presents a robust data analysis framework to deconstruct consumer preference using a dataset where 118 consumers rated their liking of 27 black drip coffee samples, the adequacy of select attributes on just-about-right (JAR) scales, and the sensory profile of the coffees with a check-all-that-apply (CATA) task. We integrated four feature-ranking methods to identify key sensory drivers, which informed the development of predictive models to forecast consumer liking. A novel consumer segmentation technique was also introduced, applying k-Means clustering to consumers’ individual preference correlation vectors. JAR acidity, JAR flavor intensity, and CATA sweetness were found to be primary drivers of liking across the population (p-value < 1e-70). The resulting predictive models demonstrated strong performance even with a limited set of 3 sensory features. Consumers were clustered into two segments with contrasting preferences for 12 different sensory attributes. The proposed analytical pipeline provides a comprehensive approach to sensory and consumer data, enabling both the prediction of general consumer liking and the identification of distinct preference segments.

An important goal for personalized diet systems is to improve nutritional quality without compromising convenience or affordability. We present an end-to-end framework that converts dietary standards into complete meals with minimal change. Using the What We Eat in America (WWEIA) intake data for 135,491 meals, we identify 34 interpretable meal archetypes that we then use to condition a generative model and a portion predictor to meet USDA nutritional targets. In comparisons within archetypes, generated meals are better at following recommended daily intake (RDI) targets by 47.0%, while remaining compositionally close to real meals. Our results show that by allowing one to three food substitutions, we were able to create meals that were 10% more nutritious, while reducing costs 19-32%, on average. By turning dietary guidelines into realistic, budget-aware meals and simple swaps, this framework can underpin clinical decision support, public-health programs, and consumer apps that deliver scalable, equitable improvements in everyday nutrition.

Modern nutrition science still lacks a comprehensive, machine-readable map linking diet to molecular composition and biological effects. Here we present FoodAtlas, a large-scale knowledge graph that links 1430 foods to 3610 chemicals, 2181 diseases, and 958 flavor descriptors through 96,981 provenance-tracked edges. A transformer-based text-mining pipeline extracted 48,474 quantitative food–chemical associations from 125,723 literature sentences ( F ₁ = 0.67) and integrated them with 23,211 chemical–disease assertions from the Comparative Toxicogenomics Database, 15,222 chemical-bioactivity records from ChEMBL, 3645 flavor annotations from FlavorDB and PubChem, and 6429 taxonomic relationships. Graph embeddings revealed six dietary modules whose signature metabolites delineate distinct, multisystem disease-risk trajectories. Models built on FoodAtlas demonstrate practical utility: a bioactivity predictor achieved strong correlation with antioxidant assays ( R ² = 0.52; ρ = 0.72), and a substitution engine reduced simulated total disease risk by 11.9%.