Multi-sample, multi-platform isoform quantification using empirical Bayes

Abstract

Accurate quantification of RNA isoform abundance is crucial for understanding gene regulation, cellular behavior, and disease mechanisms. While short-read (SR) sequencing provides high-throughput and cost-effective transcript quantification, it suffers from read-to-transcript ambiguity. Long-read (LR) sequencing reduces this ambiguity but faces challenges such as high error rates, biases, and lower throughput. Existing methods rely on either SR or LR data and operate on single or merged samples, failing to leverage the variability across multiple samples and the complementary strengths of both technologies. As a result, they struggle to accurately quantify low-abundance and moderate-expressed isoforms and often require complex models for sample-specific bias correction. To address these limitations, we introduce JOLI, a hierarchical model that leverages multi-sample learning to enhance transcript quantification by jointly integrating SR and LR sequencing data. By incorporating multi-sample learning, JOLI captures shared transcript structures, corrects for systematic biases, and enhances statistical power, particularly for low- and moderate-abundance isoforms. Our model applies an empirical Bayes framework, learning a shared prior across samples to improve inference consistency. By jointly modeling SR and LR data, it integrates the strengths of both technologies, achieving higher accuracy and reproducibility in transcript quantification. Through benchmarking on simulated and real RNA-seq datasets, we show that JOLI consistently outperforms single-sample EM method by improving ranking consistency, proportional agreement, and estimation accuracy while enhancing reproducibility. Specifically, in simulations, JOLI multi-sample improves Spearman correlation by 9.8% for LR and 7.7% for SR data compared to single-sample method, while for real data, the improvements are 2.56% (LR) and 1.28% (SR), respectively. Multi-sample learning further improves the quantification of isoforms with low to moderate expression levels. Furthermore, JOLI performs competitively with state-of-the-art methods, highlighting its robustness in transcript quantification.