Introduction
Understanding the temporal dynamics of the human immune response to viral pathogens is essential for advancing vaccine development and pandemic preparedness. However, researchers often encounter significant obstacles when utilizing existing data. Natural infection studies frequently lack pre-exposure baselines, and controlled trial data is often fragmented across disparate repositories like GEO, ImmPort, and ArrayExpress. Furthermore, inconsistent metadata and technical artifacts—such as gene name corruption from spreadsheet software—hinder the large-scale application of machine learning.
To address these challenges, researchers have introduced HR-VILAGE-3K3M, a curated, "AI-ready" longitudinal transcriptomic dataset. By harmonizing 66 studies, this resource provides a unified foundation for training models and identifying early biomarkers for diseases including Influenza and COVID-19.
Methods and Pipeline
The construction of HR-VILAGE-3K3M involved a rigorous pipeline designed to transform raw, fragmented data into a high-quality, harmonized repository.
Key methodological steps included:
- Standardization: Gene identifiers were aligned to official HGNC symbols, correcting common errors like Excel-induced date conversions (e.g., "SEPTIN14" becoming "14-Sep").
- Unified Preprocessing: Consistent normalization techniques were applied—RMA for microarrays and log-transformation for RNA-seq—alongside stringent quality control to ensure cross-study comparability.
- Metadata Harmonization: Outcome definitions, particularly vaccine response based on antibody titers, were standardized across cohorts.
- Benchmarking: The dataset's utility was validated by testing various batch-effect correction methods and predictive architectures, ranging from Logistic Regression to Transformers.
Dataset Characteristics
The HR-VILAGE-3K3M repository represents a significant scale-up in available immunological data, integrating 66 studies involving 3,178 subjects. It encompasses 14,136 gene expression profiles and over 2.56 million single cells.
The dataset covers a wide range of pathogens, including Influenza (H1N1, H3N2), SARS-CoV-2, RSV, and HRV, with samples sourced from diverse tissues such as whole blood, PBMCs, and nasal swabs.
Key Insights and Benchmarking Results
The primary value of HR-VILAGE-3K3M lies in its "AI-readiness," lowering the barrier for machine learning researchers to engage with complex biological data. Benchmarking revealed several critical insights:
- Batch Correction Trade-offs: While methods like ComBat effectively remove technical noise, they risk suppressing genuine biological signals, such as the variance between vaccine responders and non-responders.
- Model Performance: A PCA-based Logistic Regression model surprisingly outperformed several deep learning architectures in initial benchmarks. However, Transformers demonstrated superior potential in capturing the intricate temporal dynamics of immune responses.
- Predictive Accuracy: The optimized pipeline achieved an F1 score of 0.822 and an AUC of 0.748 in predicting vaccine responsiveness.
- Biological Validation: The dataset successfully captured expected temporal shifts, such as the peak in B cell proportions and transcriptional activity between Day 1 and Day 7 post-vaccination.
Limitations and Future Directions
While HR-VILAGE-3K3M is a robust resource, it is currently limited to transcriptomic data. Future iterations aim to incorporate multi-omics layers, such as proteomics and genetic variants (SNPs), to provide a more holistic view of systems immunity. Additionally, efforts continue to recover missing subject-level antibody data for certain cohorts.
Ultimately, HR-VILAGE-3K3M serves as a critical stepping stone toward developing foundation models of the human immune system, which can be fine-tuned for tasks like predicting the severity of emerging viral threats. The dataset is publicly available on Hugging Face.
Summary updated