Privacy preserving validation for multiomic prediction models

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Research

Background:

• Reproducibility of ribonucleic acid (RNA) data analysis across different laboratories is a significant challenge in cancer research.
• Variations in RNA library preparation and quantification hinder the application and validation of molecular predictors on new datasets.
• Current RNA correction methods often require access to sensitive patient-level data, limiting data sharing and predictor validation.

Purpose of the Study:

• To introduce SpinAdapt, an unsupervised RNA correction algorithm designed to facilitate the transfer of molecular models without necessitating patient-level data access.
• To address the critical need for enhanced research reproducibility and patient data privacy in multi-dataset RNA analyses.
• To develop a method that enables unbiased evaluation of predictors on validation cohorts.

Main Methods:

• SpinAdapt employs an unsupervised approach to compute data corrections using only aggregate statistics from each dataset.
• This method ensures patient data privacy by avoiding direct access to individual-level information.
• The algorithm's performance is evaluated against established methods like Seurat and ComBat using publicly available cancer datasets.

Main Results:

• SpinAdapt demonstrates superior performance compared to existing RNA correction methods, including Seurat and ComBat, on TCGA and ICGC cancer studies.
• The algorithm successfully corrects new samples, allowing for unbiased evaluation on independent validation cohorts.
• Despite the inherent privacy-performance trade-off, SpinAdapt achieves robust results while preserving patient data confidentiality.

Conclusions:

• SpinAdapt offers a novel paradigm for RNA data correction, significantly enhancing research reproducibility in cancer studies.
• The algorithm effectively enables the transfer of molecular models across datasets while upholding stringent patient data privacy standards.
• This approach is expected to accelerate the validation and clinical translation of molecular predictors in cancer research.