Privacy Preserving RNA-Model Validation Across Laboratories

Talal Ahmed, Mark Carty, Stéphane Wenric, Jonathan R. Dry, Ameen A. Salahudeen, Aly A. Khan, Eric Lefkofsky, Martin C. Stumpe, Raphael PelossofbioRxiv 2021 · DOI

Abstract

Reproducibility of results obtained using RNA data across labs remains a major hurdle in cancer research. Often, molecular predictors trained on one dataset cannot be applied to another due to differences in RNA library preparation and quantification. While current RNA correction algorithms may overcome these differences, they require access to all patient-level data, which necessitates the sharing of training data for predictors when sharing predictors. Here, we describe SpinAdapt, an unsupervised RNA correction algorithm that enables the transfer of molecular models without requiring access to patient-level data. It computes data corrections only via aggregate statistics of each dataset, thereby maintaining patient data privacy. Furthermore, SpinAdapt can correct new samples, thereby enabling evaluation of validation cohorts. Despite an inherent tradeoff between privacy and performance, SpinAdapt outperforms current correction methods that require patient-level data access. We expect this novel correction paradigm to enhance research reproducibility and patient privacy. Finally, SpinAdapt lays a mathematical framework that can be extended to other -omics modalities.

Introduction

In the field of oncology, the development of molecular predictors—tools that identify tumor subtypes or predict patient survival—holds immense promise for personalized medicine. However, a significant barrier persists: reproducibility. A model trained to high accuracy in one institution often fails when applied to data from another. This failure is primarily driven by "batch effects," which are technical variations introduced by different sequencing platforms, library preparation protocols, or even environmental conditions during processing.

To mitigate these effects, researchers typically employ batch correction algorithms. However, standard methodologies such as ComBat or Seurat require direct access to raw, patient-level data from both the source and target institutions. In an era of stringent data protection regulations like GDPR and HIPAA, sharing raw genomic data is a complex legal and ethical challenge. This creates data "silos," preventing the global validation of life-saving diagnostic models.

The SpinAdapt Framework

SpinAdapt is an unsupervised RNA correction algorithm designed to bypass the need for raw data sharing through a privacy-preserving "fit-transform" paradigm. Instead of requiring individual patient profiles, SpinAdapt operates on aggregate statistics, or "data factors."

Core Mechanism

The algorithm utilizes the Principal Component Analysis (PCA) basis, gene-wise means, and standard deviations of a dataset. These statistics capture the essential biological structure and variance of the cohort without exposing sensitive individual information. The core of SpinAdapt is a non-convex optimization problem that learns a linear transformation matrix, $A$. This matrix aligns the PCA basis of the "source" laboratory with that of the "target" laboratory.

To ensure the model remains robust and avoids overfitting to technical noise, the authors employ projected gradient descent with a matrix-rank constraint. This mathematical framework allows the model to find the optimal alignment in a low-rank latent space.

Inductive vs. Transductive Learning

A critical distinction of SpinAdapt is its inductive nature. Most traditional batch correction methods are transductive, meaning they must process the entire batch of data simultaneously to perform correction. SpinAdapt, however, creates a reusable "correction rule." Once the transformation is learned, it can be applied to new, "held-out" samples individually. This is vital for clinical applications where a model must process a single new patient sample without re-running the entire pipeline.

Experimental Validation

The researchers validated SpinAdapt using a massive meta-dataset of 4,076 samples across three technological platforms (RNA-Seq and two microarray formats). The study covered several cancer types, including bladder, breast (TCGA-BRCA), colorectal, and pancreatic cancers.

To establish a "ground truth," the authors utilized a unique cohort of 481 breast cancer patients whose samples were profiled using both RNA-seq and microarrays. This allowed for a direct comparison of how well the algorithm could harmonize data across different technologies.

Key Results and Insights

The study yielded several surprising results that challenge the conventional wisdom regarding the tradeoff between privacy and performance.

Superior Diagnostic and Prognostic Accuracy

Despite not having access to raw patient data, SpinAdapt frequently outperformed "full-access" methods:

  • Subtype Classification: In tumor subtype classification tasks, SpinAdapt significantly outperformed Seurat3 in 7 out of 17 subtypes and ComBat in 11 out of 17 subtypes.
  • Survival Prediction: When transferring Cox Proportional Hazards models, SpinAdapt improved the C-index (a measure of predictive accuracy) significantly. In colorectal cancer, the C-index rose from a baseline of 0.52 to 0.63 after correction.
  • Data Integration: Using Silhouette scores to measure how well different batches mixed while maintaining biological distinctions, SpinAdapt provided superior integration for breast and colorectal cancer datasets compared to existing methods.

The Robustness of the PCA Basis

The success of SpinAdapt suggests that the PCA basis of a transcriptomic dataset is a powerful representation. It effectively captures the latent biological signals—such as tumor subtypes—while acting as a natural mask for individual privacy. By aligning these latent spaces rather than individual data points, SpinAdapt achieves a more robust correction that is less sensitive to outliers.

Limitations and Future Directions

While SpinAdapt represents a significant leap forward, the authors note a few limitations. There remains an inherent tradeoff between privacy and performance; in specific instances, such as bladder cancer integration, non-private methods like Scanorama still maintained a slight performance edge.

However, the mathematical foundation of SpinAdapt is highly extensible. While the current iteration relies on PCA, the framework could be adapted for other matrix factorization techniques like Non-negative Matrix Factorization (NMF) or Zero-Inflated Factor Analysis (ZIFA), which may be better suited for sparse single-cell data. Furthermore, the authors envision extending this paradigm to other -omics modalities, including proteomics, methylomics, and copy number variation (CNV) data.

Conclusion

SpinAdapt demonstrates that privacy-preserving methods can meet, and even exceed, the performance of traditional data-sharing approaches. By enabling the validation of molecular models across institutional boundaries without compromising patient confidentiality, SpinAdapt provides a scalable solution to the reproducibility crisis in genomic research.

Summary updated

Abstract

Abstract Reproducibility of results obtained using RNA data across labs remains a major hurdle in cancer research. Often, molecular predictors trained on one dataset cannot be applied to another due to differences in RNA library preparation and quantification. While current RNA correction algorithms may overcome these differences, they require access to all patient-level data, which necessitates the sharing of training data for predictors when sharing predictors. Here, we describe SpinAdapt, an unsupervised RNA correction algorithm that enables the transfer of molecular models without requiring access to patient-level data. It computes data corrections only via aggregate statistics of each dataset, thereby maintaining patient data privacy. Furthermore, SpinAdapt can correct new samples, thereby enabling evaluation of validation cohorts. Despite an inherent tradeoff between privacy and performance, SpinAdapt outperforms current correction methods that require patient-level data access. We expect this novel correction paradigm to enhance research reproducibility and patient privacy. Finally, SpinAdapt lays a mathematical framework that can be extended to other -omics modalities.
BibTeX
@misc{ahmed2021privacy,
  author = {Talal Ahmed and Mark Carty and Stéphane Wenric and Jonathan R. Dry and Ameen A. Salahudeen and Aly A. Khan and Eric Lefkofsky and Martin C. Stumpe and Raphael Pelossof},
  title = {{Privacy Preserving RNA-Model Validation Across Laboratories}},
  year = {2021},
  doi = {10.1101/2021.04.01.437893},
  howpublished = {bioRxiv},
}
Figure