Personalized Plant-Based Protein Formulation-II
Description
The framework is designed to work with a multi-modal dataset that captures per-user or per-sample biological, clinical, behavioral, and nutritional information. Each sample (e.g., one individual or one formulation scenario) consists of up to four raw data streams plus a derived structural input. The dataset can be real (from the provided dataset/ folder or external cohorts) or synthetic (for development and benchmarking); the model architecture and training pipeline support both. Provided real dataset (dataset/ folder) The repository includes a real multi-modal dataset suitable for training and inference when not using synthetic data: demographic.csv: Subject demographics and survey data (e.g. NHANES-style). Columns include subject ID (SEQN), age (RIDAGEYR), sex (RIAGENDR), race/ ethnicity, education, income, household size, and sampling weights. One row per subject (~10k+ rows). Used to supply or complement biomarker features (e.g. subset of 15 clinical/demographic variables per sample). diet.csv: Dietary recall data keyed by SEQN. Columns include total calories (DR1TKCAL), protein (DR1TPROT), carbohydrates, fats, fiber, sugars, vitamins, and other nutrients per recall day. Multiple rows per subject can be used as time steps to form longitudinal dietary sequences (e.g. select 12 features, pad/truncate to max_len 64). labs.csv: Clinical lab values keyed by SEQN (blood, urine, lipids, glucose, hemoglobin, etc.). Used as the primary source for clinical biomarker features; a subset of 15 columns (or derived features) is passed to load_biomarkers() after optional joining with demographic.csv. Additional files: examination.csv, questionnaire.csv, medications.csv; daily_food_nutrition_dataset.csv (food-level calories, protein, carbs, etc.); FOOD-DATA-GROUP1–5.csv and diet_recommendations_dataset.csv; diet-type CSVs (vegan, keto, paleo, mediterranean, dash, All_Diets). These can support dietary sequence construction or nutritional knowledge (e.g. per-source EAA/digestibility) with appropriate preprocessing. Microbiome: No gut microbiome CSV is provided in dataset/. Microbiome input is either loaded from an external file (path or DataFrame passed to load_microbiome_profile) or generated synthetically when using synthetic mode or training on generate_synthetic_batch(). Samples are linked by SEQN across demographic, diet, and labs; train/validation splits are by subject (e.g. 80/20) to avoid leakage. The app (app.py) accepts uploaded CSV/JSON for microbiome, biomarkers, and dietary sequence when not using synthetic data.
Files
Steps to reproduce
Dataset size (row counts) Row counts per file in the dataset/ folder (data rows only; header excluded): File Rows All_Diets.csv 7,806 dash.csv 1,745 demographic.csv 10,175 diet.csv 9,813 diet_recommendations_dataset.csv 1,000 examination.csv 9,813 keto.csv 1,512 labs.csv 9,813 mediterranean.csv 1,753 paleo.csv 1,274 vegan.csv 1,522 daily_food_nutrition_dataset.csv 651 FOOD-DATA-GROUP1.csv 551 FOOD-DATA-GROUP2.csv 319 FOOD-DATA-GROUP3.csv 571 FOOD-DATA-GROUP4.csv 232 FOOD-DATA-GROUP5.csv 722 questionnaire.csv 10,175 medications.csv 20,194 METADATA/Combined_FOOD_METADATA.csv 25 METADATA/GROUP-1_analysis_summary.csv 5 METADATA/GROUP-2_analysis_summary.csv 5 METADATA/GROUP-3_analysis_summary.csv 5 METADATA/GROUP-4_analysis_summary.csv 5 METADATA/GROUP-5_analysis_summary.csv 5