🧬 biology

HR-VILAGE-3K3M: A Human Respiratory Viral Immunization Longitudinal Gene Expression Dataset for Systems Immunity

The paper introduces HR-VILAGE-3K3M, a comprehensive and AI-ready repository that harmonizes longitudinal transcriptomic data from 3,178 subjects across 66 studies to facilitate the discovery of immune mechanisms and biomarkers for respiratory viral infections.

Original authors: Xuejun Sun, Yiran Song, Xiaochen Zhou, Ruilie Cai, Yu Zhang, Xinyi Li, Rui Peng, Jialiu Xie, Yuanyuan Yan, Muyao Tang, Prem Lakshmanane, Baiming Zou, James S. Hagood, Raymond J. Pickles, Didong Li, Fe

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xuejun Sun, Yiran Song, Xiaochen Zhou, Ruilie Cai, Yu Zhang, Xinyi Li, Rui Peng, Jialiu Xie, Yuanyuan Yan, Muyao Tang, Prem Lakshmanane, Baiming Zou, James S. Hagood, Raymond J. Pickles, Didong Li, Fei Zou, Xiaojing Zheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to understand how the human body fights off a cold or the flu. Scientists have been running hundreds of experiments over the years, measuring the "instruction manuals" (genes) inside our cells to see how they change when we get vaccinated or exposed to a virus.

The problem is that these experiments are scattered all over the place. They are stored in different digital warehouses, written in different languages, and often missing key details like "who was this person?" or "how did they feel?" It's like trying to solve a giant jigsaw puzzle where the pieces are from 66 different boxes, some are upside down, and the picture on the box is missing.

Enter HR-VILAGE-3K3M.

Think of this new dataset as a massive, super-organized library that the authors built to fix this mess. Here is what they did, using simple analogies:

1. The Great Cleanup (Data Curation)

The researchers gathered data from 3,178 people across 66 different studies.

The Mess: Some studies used old technology (like a flip phone), others used new tech (like a smartphone). Some wrote gene names like "Septin 14," while others accidentally wrote "14-Sep" because their computer spreadsheet got confused. Some files were missing the "who" and "what."
The Fix: The team acted like digital librarians and translators. They:
- Fixed the spelling errors in gene names.
- Translated different file formats so they all speak the same language.
- Called the original scientists to ask, "Hey, we have your data, but we're missing the antibody numbers. Can you send them?"
- Removed duplicate entries (like finding the same person's photo twice in the album).

2. The Time Machine (Longitudinal Data)

Most medical studies just take one snapshot: "Here is your blood today." But this library is special because it's a movie, not a photo.

They have data collected from the same people over and over again—sometimes daily, sometimes over months.
This lets scientists watch the immune system evolve in real-time. They can see exactly when the body's defenses wake up after a vaccine or when a virus starts to take over.

3. What's Inside the Library?

The library contains two main types of "movies":

The Crowd Shot (Bulk Data): This looks at the whole blood sample at once, like taking a photo of a stadium crowd. It tells you the average activity of everyone there.
The Close-Up (Single-Cell Data): This zooms in on individual cells, like taking a photo of every single fan in the stadium. It shows exactly which specific cells are doing the work.

4. Proving It Works (The "Test Drive")

To make sure their library is actually useful, the authors ran three "test drives":

The ID Check: They asked the computer to guess a person's age and gender just by looking at their gene data. The computer was almost perfect (99% accurate for gender, very high for age), proving the data is clean and reliable.
The Vaccine Predictor: They tried to predict who would have a strong immune response to a flu shot. They tested different math tools to see which one worked best. They found that a specific method called "Quantile Normalization" (a way of smoothing out the data differences) helped the computer make the best guesses.
The Cell Detective: They compared the "Crowd Shot" (bulk) with the "Close-Up" (single-cell) data. They found that both methods agreed on how the immune cells changed over time, confirming that the library is consistent.

5. Why This Matters (According to the Paper)

The authors say this library is a benchmark (a gold standard for testing).

It helps scientists build better Artificial Intelligence (AI) models. Just like you need a huge library of books to train a smart robot to read, AI needs huge, clean datasets to learn how the immune system works.
It allows researchers to test new math tools to fix data errors or predict disease outcomes.
It helps discover biomarkers (early warning signs) of how our bodies react to viruses.

In short: The authors took a chaotic pile of 66 different scientific studies, cleaned them up, organized them, and put them in one giant, easy-to-use digital library. This allows other scientists to skip the boring cleanup work and immediately start using AI and math to understand how our bodies fight respiratory viruses.

Technical Summary: HR-VILAGE-3K3M

Problem Statement
Respiratory viral infections pose a significant global health burden, yet the cellular immune mechanisms underlying protection and pathology remain incompletely understood. While natural infection cohorts provide real-world data, they often lack pre-exposure baselines and controlled timing. Conversely, human inoculation and vaccination trials generate well-structured longitudinal transcriptomic data; however, these datasets are currently scattered across public repositories (e.g., GEO, ImmPort, ArrayExpress) and processed inconsistently.

Key barriers to integrative and AI-driven analysis include:

Data Readiness: Public repositories host data at varying levels of processing (raw, partially processed, normalized), often generated by distinct study-specific pipelines, leading to severe batch effects.
Metadata Fragmentation: Critical phenotypic data (infection status, symptoms, antibody titers) are often fragmented across supplementary files or different repositories, with sample IDs frequently mismatched between expression and metadata files.
Technical Inconsistencies: Data are profiled on diverse platforms (microarray, bulk RNA-seq, scRNA-seq) using inconsistent gene symbols and identifiers.
Quality Control Gaps: Many datasets lack rigorous preprocessing evaluation, containing technical artifacts (e.g., negative values, discrete expression levels) or duplicate samples from the same subjects deposited under different studies.

Methodology
The authors developed HR-VILAGE-3K3M (Human Respiratory Viral Immunization LongitudinAl Gene Expres- sion), a curated benchmark repository integrating data from 66 studies involving 3,178 subjects. The construction workflow involved:

Data Collection: Aggregation of gene expression profiles, metadata, and raw antibody data from GEO, ImmPort, and ArrayExpress. This included targeted outreach to study authors to retrieve missing subject-level antibody titers.
Standardized Preprocessing:
- Bulk Data: Microarray data (Affymetrix and Illumina) were processed using RMA normalization, probe-to-gene mapping, Quantile Normalization (QN), and log2 transformation. RNA-seq data were mapped to HGNC gene symbols, with raw counts provided or generated from FASTQ files via alignment and quantification.
- Single-Cell Data: Data in various formats (h5ad, 10x matrix) were integrated into Python AnnData objects, filtered (genes expressed in <10 cells removed), and normalized (total-count normalization to 10,000, log1p transformation).
Harmonization and QC:
- Gene Symbol Standardization: All symbols were aligned to official HGNC nomenclature using the HGNChelper package to correct Excel-induced formatting errors.
- Duplicate Removal: Duplicate samples from the same subjects across studies were identified via Spearman correlation of expression profiles and removed, retaining the version with the most comprehensive timepoint coverage.
- Metadata Unification: Sample IDs were aligned with expression matrices, and subject-level identifiers were standardized to link antibody measurements across studies.
- Outcome Definition: Infection and symptom statuses were standardized. For influenza vaccines, response status was defined by Hemagglutination Inhibition (HAI) titers (High: MFC ≥4, Day 28 ≥40; Non-responders: MFC ≤1).
Benchmark Analyses: The authors validated the dataset through three analyses:
- Biological Trait Prediction: Sex and age prediction using baseline whole-blood transcriptomes to assess data quality and platform consistency.
- Outcome Prediction: Predicting antibody responder status in influenza vaccine studies using longitudinal bulk data, comparing batch correction methods (QN, ComBat, Regression) and modeling strategies (PCA-GEE, RNN, LSTM, GRU, Transformer).
- Cell Type Annotation: Comparing immune cell proportions derived from paired bulk and single-cell SARS-CoV-2 vaccine data using xCell and automated annotation tools (SingleR, GPTCelltype).

Key Results

Dataset Composition: HR-VILAGE-3K3M contains 14,136 expression profiles from 3,178 participants. It includes 15 inoculation studies (5 viruses: H3N2, H1N1, HRV, RSV, SARS-CoV-2), 47 vaccination studies (Influenza, COVID-19), and 4 mixed exposure studies. Platforms include microarray (52%), RNA-seq (38%), and 10x Genomics scRNA-seq (11%).
Quality Control Validation:
- Sex Prediction: Y-chromosome marker-based classifiers achieved near-perfect discrimination (ROC-AUC: 0.989 for microarray, 0.991 for RNA-seq), confirming sample labeling accuracy.
- Age Prediction: Models achieved strong predictive performance ( $R^2 \approx 0.83$ for microarray, $0.71$ for RNA-seq using Ridge regression), demonstrating that baseline transcriptomes encode substantial age-related variation.
Batch Effect and Predictive Modeling:
- Batch Correction: Quantile Normalization (QN) preserved biological signal (responder status separation) better than ComBat, which over-corrected and reduced outcome separation.
- Prediction Performance: The Transformer model with QN correction achieved the highest performance in predicting antibody responders (AUC: 0.737, Accuracy: 0.663), outperforming recurrent neural networks (RNN, LSTM, GRU) and other correction methods.
Multi-Modal Consistency: Analysis of paired bulk and scRNA-seq data from SARS-CoV-2 vaccination showed consistent temporal trends (e.g., B cell increase, Monocyte/NK decrease from Day 1 to Day 7) between scRNA-seq proportions and bulk xCell enrichment scores.

Significance and Claims
The paper positions HR-VILAGE-3K3M as a foundational resource for systems immunology and AI-driven research. Its primary contributions are:

Accessibility: It transforms fragmented, raw, and inconsistent public data into a harmonized, ready-to-use resource, bypassing the time-consuming steps of data discovery, cleaning, and integration.
Benchmarking: It serves as a rigorous benchmark for developing and evaluating computational methodologies, including batch effect correction, gene expression imputation, deep learning prediction, and deconvolution of bulk data using single-cell references.
AI Enablement: The scale and heterogeneity of the dataset support the pretraining of foundation models to capture complex immune dynamics, facilitating transfer learning for smaller, underpowered datasets.
Scientific Utility: The resource enables the discovery of early biomarkers, identification of causal genes, and the study of temporal immune dynamics across diverse pathogens and exposure types (vaccination vs. inoculation).

The authors emphasize that the dataset is publicly available on Hugging Face with analysis code on GitHub, designed to accelerate reproducible and scalable research in vaccine and antiviral development.