A computational framework to improve cross-platform implementation of transcriptomics signatures

Abstract

The emergence of next-generation sequencing technologies and computational advances have expanded our understanding of gene expression regulation (i.e., the transcriptome). This has also led to an increased interest in using transcriptomic biomarkers to improve disease diagnosis and stratification, to assess prognosis and predict the response to treatment. Significant progress in identifying transcriptomic signatures for various clinical needs has been made, with large discovery studies accounting for challenges such as patient variability, unwanted batch effects, and data complexities; however, obstacles related to the technical aspects of cross-platform implementation still hinder the successful integration of transcriptomic technologies into standard diagnostic workflows. In this article, we discuss the challenges associated with integrating transcriptomic signatures derived using high-throughput technologies (such as RNA-sequencing) into clinical diagnostic tools using nucleic acid amplification (NAA) techniques. The novelty of the proposed approach lies in our aim to embed constraints related to cross-platform implementation in the process of signature discovery. These constraints could include technical limitations of amplification platform and chemistry, the maximal number of targets imposed by the chosen multiplexing strategy, and the genomic context of identified RNA biomarkers. Finally, we propose to build a computational framework that would integrate these constraints in combination with existing statistical and machine learning models used for signature identification. We envision that this could accelerate the integration of RNA signatures discovered by high-throughput technologies into NAA-based approaches suitable for clinical applications.

Keywords: Diagnostics; Host-response; Molecular test; Multiplex PCR; Nucleic acid amplification techniques; PCR-based technologies; RNA sequencing; Transcriptomic signatures.

Fig. 1

Embedding constraints related to cross-platform implementation in the feature selection process to improve the translation of transcriptomics signatures into diagnostic tools. In this figure, we depict classical methodologies for gene signature discovery and implementation (blue background), and propose novel strategies to ensure that signatures retain optimal classification accuracy after cross-platform transfer (red background). Classically, the down-selection of a large set of differentially expressed genes in a final sparse signature is mainly based on statistical criteria and machine learning methods (top blue panel). Here, we propose to enhance this feature selection process by designing algorithms that would further incorporate constraints associated with the target implementation platform and chemistry. In the bottom left panel, we highlight how three exons with different characteristics (differential expression in cases vs. controls and length) may be suited for quantification using real-time LAMP (qLAMP), reverse transcription real-time PCR (RT-qPCR) and digital droplet PCR (ddPCR). In the bottom middle panel, we present how different multiplexing strategies might impact the total number of transcripts included in the final sparse signature. In the bottom right panel, we show that different assay designs (i.e., the location of primers on the genomic region of interest) might impact the way splicing variants are quantified on the discovery and implementation platform (RNA-Seq vs. RT-qPCR). Taken together, we suggested embedding these criteria upstream of the discovery process to ensure successful implementation of transcriptomics signatures into validated diagnostic tools. NAAT: nucleic acid amplification technique, AUC: area under the receiver operating characteristic curve, FC: fold-change, qLAMP: quantitative loop-mediated isothermal amplification, RT-qPCR: reverse transcription real-time polymerase chain reaction (PCR), ddPCR: digital droplet PCR. Figure created with BioRender.