Integrating Traditional AI/ML, LLMs, and Quantum Approaches

To capture the complexity of obesity-related biological mechanisms, we integrated multiple modeling paradigms—from classical AI/ML pipelines to emerging quantum computing frameworks. Each approach complements the others: traditional models ensure interpretability, while LLMs and AI agents enhance adaptability, and quantum simulations explore molecular-scale interactions.

Comparison of Traditional AI/ML, Generative AI/LLM, AI Agents, and Quantum–Mathematical Models

Category Traditional AI / ML Generative AI / LLM AI Agents Quantum–Mathematical / System Models
Definition Systems that analyze structured data for prediction and classification. Transformer-based AI that understands unstructured data and generates text or code. Multi-agent systems that orchestrate ML, LLM, APIs, and workflows to automate complex tasks. Models that combine physics-based equations (ODEs) and quantum simulations to represent biological and behavioral processes mechanistically.
Core Features Regression, classification, clustering, and pattern recognition. Text generation, summarization, and semantic understanding. Coordination, feedback loops, and decision orchestration. Simulation of metabolic systems via differential equations and probabilistic modeling (e.g., AMPK pathway, calorie balance).
Input Data Types Numeric datasets (calories, BMI, heart rate, etc.). Textual and conversational data (user input, feedback). Multimodal data (sensor logs, EHR, app activity). Physiological constants, empirical parameters, and biological kinetics.
Strengths Efficient at large-scale structured analysis. Natural-language reasoning and adaptive consultation. Automates repetitive coaching and decision support. Enables hypothesis testing, parameter sensitivity, and pathway-level prediction for microbiome and metabolism.
Limitations Weak in contextual reasoning and personalization. Risk of hallucination and factual inaccuracy. Requires complex governance and compliance tracking. Dependent on data calibration and biological validation; limited in real-time adaptability.
Use Cases Predicting calorie expenditure, clustering user risk profiles. Personalized plan recommendations, AI chatbot for adherence. Coach agent for daily guidance and plan orchestration. ODE-based weight prediction model, exosome pathway modeling, and exploratory quantum docking simulation.
Recommendation Use for structured predictions (numerical health metrics). Use for communication, summarization, and adaptive feedback. Use for coordination and personalization logic. Model complex interactions beyond classical limits to understand underlying biological or metabolic mechanisms.

Classical Modeling: Weight Management ODE Framework

We first established a deterministic mathematical model describing how energy balance drives weight change.

At the foundation, we built a differential equation–based model representing changes in energy expenditure and caloric intake over time. The model integrates variables for physical activity, microbiome influence, and lipid metabolism to estimate dynamic weight trends under different lifestyle or dietary interventions.

This classical foundation enables controlled simulation of how microbiome-derived exosomes could modulate energy balance through AMPK pathway activation or lipid-storage inhibition.

The fundamental principle behind weight regulation is straightforward: when a person consumes more energy than they expend, body weight increases; conversely, when energy expenditure exceeds intake, weight decreases.

We define energy balance (EB) as the difference between energy intake (EI) and energy expenditure (EE):

EB = EI − EE

A positive energy balance indicates weight gain, while a negative balance leads to weight loss.

At any given time t, an individual’s total body weight, BW(t), can be expressed as the sum of fat mass F(t) and lean mass L(t):

BW(t) = F(t) + L(t)

The rate of change in each tissue compartment can then be represented by a system of differential equations:

$$\frac{dF}{dt} = \frac{(1-p(t))EB(t)}{\rho_F}, \frac{dL}{dt} = \frac{p(t)EB(t)}{\rho_L}$$

Here, p(t) represents the proportion of energy directed toward lean tissue change, while 1 − p(t) applies to fat tissue. The constants ρF and ρL are the energy densities of fatty and lean tissue, typically approximated as:

ρF = 9400 kcal/kg,  ρL = 1800 kcal/kg

In our project, this classical framework serves as the macroscopic foundation for modeling energy balance and weight dynamics. We extend this model by introducing biological modulation through exosome-mediated AMPK activation, a mechanism known to enhance energy expenditure and lipid oxidation.

By incorporating a time-dependent metabolic efficiency factor, α(t), which reflects microbiome-derived metabolic stimulation, the system becomes:

EB(t) = EI(t) − [EEbase(t) + α(t) ⋅ EEAMPK(t)]

This hybrid equation allows us to simulate how Lactobacillus rhamnosus exosomes could shift the body’s equilibrium toward higher energy utilization without requiring pharmaceutical intervention.

In effect, our differential model captures the biochemical impact of AMPK activation on macroscopic weight regulation, bridging molecular biology with whole-body energy modeling — a unique integration of wet-lab biotogy and computational simulation.

This foundation supports our AI-agentic simulations by providing measurable, physiology-based ground truth.

AI and Agentic Modeling: Dynamic Behavior and Personalization

To model real-world health behavior change, we developed an AI-Agent framework that translates top-level goals (e.g., “Enhance Energy Expenditure”) into atomic daily actions (e.g., “Take stairs instead of elevator”).

System Flow:

  • The AI Planner translates high-level goals into context-aware recommendations.
  • The Agent Gateway monitors user adherence and adapts plans dynamically.
  • Feedback loops ensure personalization and long-term engagement.

The agentic layer bridges data-driven personalization and biological modeling, mirroring how digital health platforms (e.g., Noom Coach) continuously learn from user feedback to improve outcomes.

We designed an agent-based decision model to translate weight management goals into structured behavioral plans. The model links Top-Level Goals, Mid-Level Plans, and Bottom-Level Atomic Actions, simulating how users interact with an AI-guided system for sustainable obesity management.

Service Layer

At the top level, users experience an intuitive Weight Management Service that merges biological insight, behavioral science, and digital health engagement. Through app-based tracking, AI coaching, and microbiome-informed dietary design, we aim to transform obesity management into a personalized, empowering, and sustainable daily experience.

Multi-model Multi-agent Unified UX

1) Rules Engine / Knowledge Graph → “Diet, Exercise, and Medication Knowledge Base”

2) Data Sources

  • User Logs → “Activity, Diet, and Sleep Data”
  • Wearables → “Steps, Heart Rate, Sleep Metrics”
  • Clinical Data → “Blood Sugar, BMI, Medication Records”

3) Application Programming Interfaces (APIs) → enable smooth data flow between literature retrieval, molecular modeling, and experimental planning components.

4) Predictive AI Model → Progress Forecasting via Continuous Feedback

5) Models

  • LLM Model → “LLM Planner for Behavior & Goal Decomposition”
  • Dynamic Model (ODE) → “Weight-Change Differential Equation Model (EI-EE)”
  • Quantum Model → “Quantum Optimization for Protein / Exosome Interaction”

Data Agentic Layer – LLM-Guided Personalization

The system employs an LLM-based agent framework to deliver intelligent, personalized engagement:

LLM Agent Manager

  • Planner Agent – Generates adaptive daily plans based on user logs.
  • Coach Agent – Provides motivational feedback and habit reinforcement.

Orchestrator

  • Personalization Agent – Adjusts recommendations according to progress, diet, and lifestyle data.

This multi-agent structure ensures that users receive continuous, context-aware coaching rather than static advice.

Model Gateway

  • Acts as the Agentic Coordination Layer — Integrating LLM, ODE, and Behavioral Models*
  • (Ensures smooth communication between user inputs, health metrics, and AI-driven predictions.)

Data Layer

Our three-tier database supports integrated insights:

  • Health DB – Stores physical activity, calorie, and sleep data.
  • Knowledge DB – Contains verified exercise and nutrition guidelines.
  • User Profile DB – Holds individual goals, health conditions (e.g., diabetes), and dietary restrictions.

The system continuously learns from these databases to refine user guidance and predict behavioral bottlenecks.

AI-guided Weight Management Framework

AI-guided Weight Management Framework

Top-Level Goals

Enhance Energy Expenditure

  • Promote physical activity, microbiome activation (AMPK pathway), and adequate sleep.

Reduce Calorie Intake

  • Manage appetite, improve food choices, and introduce structured meal timing.

Medical Support

  • Integrate clinician feedback and pharmacological guidance (e.g., GLP-1, probiotics).

Mid-Level Plans

1a. Build Exercise Routines – Create daily walking or light workout plans.

1b. Increase Daily Movement – Encourage standing, stair use, and short activity breaks during work.

1c. Improve Sleep Patterns – Implement regular sleep schedules and reduce late-night screen time.

2a. Manage Meal Plans – Plan balanced meals with appropriate macronutrients.

2b. Reduce Snacking – Substitute high-calorie snacks with fruit or yogurt alternatives.

2c. Manage Beverages – Replace sugary drinks with black coffee or water; track liquid intake.

3a. Medication Adherence – Support consistent, safe use of prescribed treatments such as GLP-1 agonists or probiotic supplements.

3b. Health Monitoring – Track key symptom data to detect early warning signs or side effects.

3c. Professional Consultation – Facilitate regular communication between users, physicians, and nutrition specialists for personalized care.

Bottom-Level Atomic Actions

Each plan is decomposed into concrete, trackable actions—the building blocks of behavioral change:

1a. Prepare workout clothes, go for a 30-minute walk, and log the activity in the app.

1b. Take the stairs, stand up every hour, and stretch briefly at your desk.

1c. Stop using your phone 30 minutes before bed and aim to sleep before 11 p.m.

2a. Plan meals each morning, choose a salad for lunch, and focus on protein at dinner.

2b. Replace chips with almonds or fruit and avoid late-night snacking.

2c. Order an Americano instead of a latte and drink 2 L of water daily (recorded in app).

3a. Take supplements or prescribed medication at the same time each day, logging intake via app reminders or a smart pill case.

3b. Record daily measurements (e.g., blood sugar, blood pressure) and sync wearable data to the Health DB for automatic alerts to the AI Coach.

3c. Schedule monthly telehealth consultations, upload recent results, and let the AI Planner generate tailored discussion topics for clinicians.

Quantum Computing

Quantum-Assisted Modeling: Exploring Molecular Interactions

Inspired by recent studies such as Cleveland Clinic & IBM (JCTC, 2024), we realized how quantum computing could dramatically accelerate early-stage drug and ingredient discovery by exploring energy landscapes unreachable by classical methods.

To push the frontier further, we explored quantum-assisted modeling for molecular interaction prediction.

While we did not perform direct quantum computation, our framework outlines how quantum optimization could accelerate exosome-protein interaction analysis in the future.

In our conceptual model, molecular binding dynamics are represented as a Quadratic Unconstrained Binary Optimization (QUBO) problem, where each possible configuration corresponds to a potential molecular state.

Then, we applied Quantum Approximate Optimization Algorithm (QAOA) to search low-energy binding states Finally, we would use hybrid AI–quantum workflows for molecular conformation prediction

Accelerating Drug Discovery & Development with Quantum

Recent research in the biopharmaceutical field highlights the growing role of quantum and hybrid modeling in protein–ligand docking, enzyme–substrate interaction, and molecular pathway optimization.

These methods are increasingly used to capture the complex energy landscapes of molecular systems that classical simulations approximate only at limited accuracy.

Drawing inspiration from these developments, our work extends the same modeling principles toward functional food and microbiome-based therapy design, rather than traditional small-molecule drug development.

This alignment reflects a broader shift toward computational–biological convergence, where data-driven and physics-informed approaches jointly advance health innovation.

Pipeline of drug discovery and development

Figure 1: Pipeline of drug discovery and development[3]
Discovery identifies candidate molecules with desirable pharmacological properties. Preclinical development evaluates safety and efficacy in animal models. Clinical development then tests safety, dosing, and effectiveness in humans through phased trials.

Quantum Computing Methods for Predicting Protein Structures

Figure 2: Quantum Computing Methods for Predicting Protein Structures[4]
Researchers at Cleveland Clinic reported in April 2024 that a quantum-computing approach produced a Zika virus NS3 helicase P-loop structure that matched experiment better (lower RMSD) than AlphaFold2 and other methods.

Quantum Exploration for Future Applications

As part of our modeling and systems research, Team Essential explored how quantum computing could play a role in the future of biological modeling and therapeutic design.

We visited the Quantum Korea Expo 2025, where we engaged with researchers and technology partners to understand emerging trends in quantum-accelerated life-science computation.

Our main insight was that while current quantum hardware is not yet ready for large-scale biological simulation, conceptual frameworks already exist that model components such as:

  • Protein folding and structure prediction using hybrid quantum–classical energy minimization.
  • Optimization of molecular binding for identifying candidate active ingredients or exosomal peptides.
  • Accelerated screening of metabolic or microbiome-derived compounds through quantum-inspired search algorithms.

These discussions shaped our long-term vision: quantum computing could one day help identify novel probiotic or peptide-based active ingredients for obesity management by simulating binding interactions at unprecedented accuracy and scale.

Although our current project remains at the conceptual stage, this exploration helped our team articulate how quantum technologies may eventually integrate into AI-assisted molecular discovery pipelines—linking fundamental computation, biological modeling, and personalized nutrition.

Quantum Korea 2025 Expo

Figure 3: Members of Team Essential, Korea, visit the Quantum Korea 2025 Expo to explore how advancements in quantum computing could support future applications of our project — particularly in modeling molecular interactions and identifying novel active ingredients for microbiome-based obesity management.

LLM-Driven Ingredient & Pathway Discovery Architecture

Finally, we designed a multi-stage AI research architecture integrating LLM-based reasoning, protein structure prediction, and quantum-informed optimization.

Three Core Stages:

  • Contextual RAG-Based Literature Search – Retrieves obesity–protein–microbiome relations from PubMed and iGEM sources.
  • AlphaFold2-Based Protein–Ligand Interaction Prediction – Simulates structure binding for candidate probiotics such as Lactobacillus.
  • Quantum Simulation-Based Complex Optimization – Conceptually maps non-linear metabolic pathways using QAOA-based quantum modeling.

Together, this pipeline represents a next-generation AI–quantum hybrid research model for precision nutrition and microbiome therapy.

By integrating advanced AI and quantum computing technologies, Team Essential aims to establish a structured and systematic research framework that enables fast and accurate identification of bioactive compounds, fostering accelerated progress in therapeutic discovery.

LLM-Driven Ingredient Discovery Architecture

Figure 4: LLM-Driven Ingredient Discovery Architecture

We built a Language Model–driven data-mining framework to accelerate ingredient selection and literature synthesis for probiotic-derived compounds.

Components:

  • Planner Agent identifies key metabolic pathways (e.g., AMPK, PPARγ).
  • Retriever Agent searches biomedical papers and databases.
  • Evaluator Agent scores each candidate based on safety, GRAS potential, and prior clinical relevance.

Model Output:

A ranked map of bioactive exosomal proteins and peptides related to lipid metabolism and energy balance.

How it informed design: guided our selection of Lactobacillus rhamnosus as the primary strain, linking its known AMPK activation to lipid metabolism improvement.

References

[1] “Weight Loss.” ACME, Brigham Young University, acme.byu.edu/00000179-d3f1-d7a6-a5fb-ffff6a200000/weightloss-pdf.

[2] “Developing Agent-Based Models of Complex Systems.” PMC, U.S. National Library of Medicine, PMC6284360, https://pmc.ncbi.nlm.nih.gov/articles/PMC6284360/.

[3] Zhou, Y., Chen, J., Cheng, J., Karemore, G., Zitnik, M., Chong, F. T., Liu, J., Fu, T., & Liang, Z. (n.d.). Quantum-machine-assisted drug discovery: Survey and perspective [Preprint].

[4] Doga, Hakan, et al. “A Perspective on Protein Structure Prediction Using Quantum Computers.” Journal of Chemical Theory and Computation, vol. 20, no. 9, 2024, pp. 3359–3378. https://doi.org/10.1021/acs.jctc.4c00067

[5] “How Quantum Computing Is Revolutionising Drug Development.” Drug Discovery World, 2025.

[6] Crooks, Gavin E. “Quantum Optimization and the Future of Molecular Modeling.” Nature Computational Science, vol. 5, no. 2, 2025, pp. 101–105.

[7] Brown, Nathan, et al. “Applications of Quantum Computing in Structural Biology and Drug Discovery.” Frontiers in Molecular Biosciences, vol. 12, 2024, p. 146233.